Specification for Free: Behavior-Driven Fuzzing with Inferred Specifications

A bug uncovered by fuzzing is seldom the product of a single input. More often, it's a family of inputs that reliably triggers the same underlying flaw.
Validating a fix means testing against that entire family, not just the one input that happened to surface the problem. In this lecture, I'll show how to
capture these families as specialized grammars that encode specific program behaviors, particularly faulty ones. You'll learn how to extract and merge such
patterns into targeted grammars, and how to use them to focus your fuzzing efforts. These grammars compose naturally with smart fuzzers, giving you fine-
grained control over input generation and making tests consistent and reproducible.

A persistent challenge in fuzzing is that leveraging input structure requires accurate specifications, which aren't always on hand. I'll show how to
sidestep this by deriving input grammars directly from the program. When source code is available, static and dynamic analysis can trace input-handling
logic from first principles. Without it, we can infer structure by observing how the program responds to inputs and iteratively refining our model through
feedback.

Together, these techniques make specification-free fuzzing practical and lay the foundation for the behavior-driven strategies we'll develop throughout the
session.

1. 1 Speci fi cations for Free:
 Behavior-Driven Fuzzing with Inferred Speci fi cations Rahul Gopinath
2. 2 Speci fi cations for Free:
 Behavior-Driven Fuzzing with Inferred Speci fi cations Rahul Gopinath
3. Prerequisites https://github.com/vrthra/summer2026-nus • Install Python 3.12 • Install Graphviz (optional) • Install Jupyter • Start Jupyter 3 $ jupyter notebook --ip='*' --NotebookApp.token='' --NotebookApp.password=''
4. Prerequisites https://github.com/vrthra/summer2026-nus • Install Python 3.12 • Install Graphviz (optional) • Install Jupyter • Start Jupyter 4 $ jupyter notebook --ip='*' --NotebookApp.token='' --NotebookApp.password=''
5. 5 Security Research | Software Verification Research
6. 6 Security Research | Software Verification Research
7. https://fuzzingbook.org 7
8. https://rahul.gopinath.org/posts 8
9. http://localhost:8888/notebooks/RoadMap.ipynb 9
10. http://localhost:8888/notebooks/RoadMap.ipynb 10
11. http://localhost:8888/notebooks/x0_0_Prerequisites.ipynb 11
12. 12 Finding hidden bugs
13. 13 Fuzzing Crash? Program Trash deck technique: 1950s - Gerald Weinberg
14. 14 $ ./fuzz [;x1-GPZ+wcckc];,N9J+?#6^6\e?]9lu
 2_%'4GX"0VUB[E/r ~fApu6b8<{%siq8Z
 h.6{V,hr?;{Ti.r3PIxMMMv6{xS^+'Hq!
 AxB"YXRS@!Kd6;wtAMefFWM(`|J_<1~o}
 z3K(CCzRH JIIvHz>_*.\>JrlU32~eGP?
 lR=bF3+;y$3lodQ<B89!5"W2fK*vE7v{'
 )KC-i,c{<[~m!]o;{.'}Gj\(X}EtYetrp
 bY@aGZ1{P!AZU7x#4(Rtn!q4nCwqol^y6
 }0|Ko=*JK~;zMKV=9Nai:wxu{J&UV#HaU
 )*BiC<),`+t*gka<W=Z.%T5WGHZpI30D<
 Pq>&]BS6R&j?#tP7iaV}-}`\?[_[Z^LBM
 PG-FKj'\xwuZ1=Q`^`5,$N$Q@[!CuRzJ2
 D|vBy!^zkhdf3C5PAkR?V((-%><hn|3='
 i2Qx]D$qs4O`1@fevnG'2\11Vf3piU37@
 5:dfd45*(7^%5ap\zIyl"'f,$ee,J4Gw:
 cgNKLie3nx9(`efSlg6#[K"@WjhZ}r[Sc
 un&sBCS,T[/3]KAeEnQ7lU)3Pn,0)G/6N
 -wyzj/MTd#A;r*(ds./df3r8Odaf?/<#r Program ✘ Random Fuzzing
15. 15 http://localhost:8888/notebooks/x1_0_GeneratingSamples.ipynb
16. http://localhost:8888/notebooks/x1_0_GeneratingSamples.ipynb 16
17. 17 http://localhost:8888/notebooks/x1_0_GeneratingSamples.ipynb
18. @app.route('/admin') def admin(): username = request.cookies.get("username") if not username: return {"Error": "Specify username in Cookie"} username = urllib.quote(os.path.basename(username)) url = "http://permissions:5000/permissions/{}".format(username) resp = requests.request(method="GET", url=url) # "superadmin\ud888" will be simpli fi ed to "superadmin" ret = ujson.loads(resp.text) if resp.status_code == 200: if "superadmin" in ret["roles"]: return {"OK": "Superadmin Access granted"} else: e = u"Access denied. User has following roles: {}".format(ret["roles"]) return {"Error": e}, 401 else:return {"Error": ret["Error"]}, 500 [ ; x 1 - G P Z + w c c k c ] ; , N 9 J + ? # 6 ^ 6 \ e ? ] 9 l u 2 _ % ' 4 G X " 0 V U B [ E / r ~ f A p u 6 b 8 < { % s i q 8 Z h . 6 { V , h r ? ; {Ti.r3PIxMMMv6{xS^+'Hq!AxB"YXRS@! Kd6;wtAMefFWM(`|J_<1~o}z3K(CCzRH J I I v H z > _ * . \ > J r l U 3 2 ~ e G P ? lR=bF3+;y$3lodQ<B89!5"W2fK*vE7v{')KC- i,c{<[~m!]o;{.'}Gj\(X}EtYetrpbY@aGZ1{P! A Z U 7 x # 4 ( R t n ! q 4 n C w q o l ^ y 6 } 0 | Ko=*JK~;zMKV=9Nai:wxu{J&UV#HaU)*Bi C < ) , ` + t * g k a < W = Z . % T 5 W G H Z p I 3 0 D < P q > & ] B S 6 R & j ? # t P 7 i a V } - } ` \ ? [ _ [ Z ^ L B M P G - FKj'\xwuZ1=Q`^`5,$N$Q@[!CuRzJ2D|vBy! ^ z k h d f 3 C 5 P A k R ? V ( ( - % > < h n | 3='i2Qx]D$qs4O`1@fevnG'2\11Vf3piU37@ 5 : d f d 4 5 * ( 7 ^ % 5 a p \ z I y l " ' f , $ee,J4Gw:cgNKLie3nx9(`efSlg6#[K"@Wjh Z}r[Scun&sBCS,T[/3]KAeEnQ7lU)3Pn,0)G/ 6N-wyzj/MTd#A;r Program 18 https://www.fuzzingbook.org/html/Fuzzer.html Traditional Fuzzing
19. 19 (CACM '90) No longer very effective
20. 20 http://localhost:8888/notebooks/x1_0_GeneratingSamples.ipynb#A-random-fuzzer.
21. • Insert Instrumentation • Generate inputs • Collect execution feedback • Branches covered during execution • Slightly Mutate Input and try again Collect inputs obtaining new coverage 21 Feedback Driven Fuzzing 21 https://www.fuzzingbook.org/html/MutationFuzzer.html
22. 22 def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene def triangle(a, b, c): __probe_enter() if a == b: __probe_1() if b == c: __probe_2() return Equilateral else: __probe_3() return Isosceles else: __probe_4() if b == c: __probe_5() return Isosceles else: __probe_6() if a == c: __probe_7() return Isosceles else: __probe_8() return Scalene def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene • Insert Instrumentation • Generate inputs • Collect execution feedback • Branches covered during execution • Slightly Mutate Input and try again Collect inputs obtaining new coverage 22 https://www.fuzzingbook.org/html/MutationFuzzer.html Feedback Driven Fuzzing
23. 23 triangle (1,1,1) def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene • Insert Instrumentation • Generate inputs • Collect execution feedback • Branches covered during execution • Slightly Mutate Input and try again Collect inputs obtaining new coverage 23 https://www.fuzzingbook.org/html/MutationFuzzer.html Feedback Driven Fuzzing
24. 24 triangle (1,1,1) def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene • Insert Instrumentation • Generate inputs • Collect execution feedback • Branches covered during execution • Slightly Mutate Input and try again Collect inputs obtaining new coverage 24 https://www.fuzzingbook.org/html/MutationFuzzer.html Feedback Driven Fuzzing
25. triangle (1,1,1) 25 triangle (1,1,2) def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene • Insert Instrumentation • Generate inputs • Collect execution feedback • Branches covered during execution • Slightly Mutate Input and try again Collect inputs obtaining new coverage Mutated 25 https://www.fuzzingbook.org/html/MutationFuzzer.html Feedback Driven Fuzzing
26. 26 triangle (1,1,3) def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene • Insert Instrumentation • Generate inputs • Collect execution feedback • Branches covered during execution • Slightly Mutate Input and try again Collect inputs obtaining new coverage Mutated 26 https://www.fuzzingbook.org/html/MutationFuzzer.html Feedback Driven Fuzzing
27. Feedback Driven Fuzzing 27 triangle (1,1,2) def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene • Insert Instrumentation • Generate inputs • Collect execution feedback • Branches covered during execution • Slightly Mutate Input and try again Collect inputs obtaining new coverage 27 https://www.fuzzingbook.org/html/MutationFuzzer.html triangle (1,1,1) triangle (1,1,3)
28. Feedback Driven Fuzzing • Insert Instrumentation • Generate inputs • Collect execution feedback • Branches covered during execution • Slightly Mutate Input and try again Collect inputs obtaining new coverage 28 triangle (1,1,2) def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene triangle (1,1,1) 28
29. 29 Minimal Framework to Track Coverage http://localhost:8888/notebooks/x1_0_GeneratingSamples.ipynb#Tracking-Coverage
30. 30 Weakness: static int is_reserved_word_token(const char *s, int len) { const char *reserved[] = { "break", "case", "catch", "continue", "debugger", "default", "delete", "do", "else", "false", "finally", "for", "function", "if", "in", "instanceof", "new", "null", "return", "switch", "this", "throw", "true", "try", "typeof", "var", "void", "while", "with", "let", "undefined", ((void *)0)}; int i; if (!mjs_is_alpha(s[0])) return 0; for (i = 0; reserved[i] != ((void *)0); i++) { if (len == (int)strlen(reserved[i]) && strncmp(s, reserved[i], len) == 0) return i + 1; } return 0; } Tokens if (x > 100) { } coverage: 20% if (x > 100) { } e coverage: 5% if (x > 100) { } el coverage: 5% if (x > 100) { } els coverage: 5% if (x > 100) { } else coverage: 25% No smooth coverage gradient in
 parsers with lexical tokens 30 Note: Constant unrolling (e.g. AFL) does not help in such lexical tokens Feedback Driven Fuzzing
31. 31 def json_raw(stm): while True: stm.skipspaces() c = stm.peek() if c == 't': return json_fixed(stm, 'true') elif c == 'f': return json_fixed(stm, 'false') elif c == 'n': return json_fixed(stm, 'null') elif c == '"': return json_string(stm) elif c == '{': return json_dict(stm) elif c == '[': return json_list(stm) elif c in NUMSTART: return json_number(stm) raise JSONError(E_MALF, stm, stm.pos) Weak points: • Need for smooth coverage gradient • Coverage only provides first level guidance 1. {"abc":[]} 2. [{"a":[]}, {"b":[]}, {"c":["ab","c"]}] 31 Feedback Driven Fuzzing
32. 32 def json_raw(stm): while True: stm.skipspaces() c = stm.peek() if c == 't': return json_fixed(stm, 'true') elif c == 'f': return json_fixed(stm, 'false') elif c == 'n': return json_fixed(stm, 'null') elif c == '"': return json_string(stm) elif c == '{': return json_dict(stm) elif c == '[': return json_list(stm) elif c in NUMSTART: return json_number(stm) raise JSONError(E_MALF, stm, stm.pos) Weak points: • Need for smooth coverage gradient • Coverage only provides first level guidance 1. {"abc":[]} 2. [{"a":[]}, {"b":[]}, {"c":["ab","c"]}] 32 Feedback Driven Fuzzing
33. 33 def json_raw(stm): while True: stm.skipspaces() c = stm.peek() if c == 't': return json_fixed(stm, 'true') elif c == 'f': return json_fixed(stm, 'false') elif c == 'n': return json_fixed(stm, 'null') elif c == '"': return json_string(stm) elif c == '{': return json_dict(stm) elif c == '[': return json_list(stm) elif c in NUMSTART: return json_number(stm) raise JSONError(E_MALF, stm, stm.pos) Weak points: • Need for smooth coverage gradient • Coverage only provides first level guidance 1. {"abc":[]} 2. [{"a":[]}, {"b":[]}, {"c":["ab","c"]}] 33 Feedback Driven Fuzzing
34. 34 JSON Parser Validation [{"id":1},{"id":2,"tags":["x","y"]}] {"user":{"name":"alice","roles":["admin","dev"]}} [{"path":["home","docs"]},{"size":{"w":1920,"h":1080}}] {"api":{"status":200,"headers":{"content-type":"json"}}} {"system":{"services":[{"name":"db"},{"name":"cache"}]}} Business Logic Where you want to reach System Internals
36. 36 Overcoming Parsers
37. 37 1y78 ✘ ( ✘ (1+2) ✔ 21214/91*293 ✔ &12133 ✘
38. 38 Fuzzing Parsers $ ./fuzz [;x1-GPZ+wcckc];,N9J+?#6^6\e?]9lu
 2_%'4GX"0VUB[E/r ~fApu6b8<{%siq8Z
 h.6{V,hr?;{Ti.r3PIxMMMv6{xS^+'Hq!
 AxB"YXRS@!Kd6;wtAMefFWM(`|J_<1~o}
 z3K(CCzRH JIIvHz>_*.\>JrlU32~eGP?
 lR=bF3+;y$3lodQ<B89!5"W2fK*vE7v{'
 )KC-i,c{<[~m!]o;{.'}Gj\(X}EtYetrp
 bY@aGZ1{P!AZU7x#4(Rtn!q4nCwqol^y6
 }0|Ko=*JK~;zMKV=9Nai:wxu{J&UV#HaU
 )*BiC<),`+t*gka<W=Z.%T5WGHZpI30D<
 Pq>&]BS6R&j?#tP7iaV}-}`\?[_[Z^LBM
 PG-FKj'\xwuZ1=Q`^`5,$N$Q@[!CuRzJ2
 D|vBy!^zkhdf3C5PAkR?V((-%><hn|3='
 i2Qx]D$qs4O`1@fevnG'2\11Vf3piU37@
 5:dfd45*(7^%5ap\zIyl"'f,$ee,J4Gw:
 cgNKLie3nx9(`efSlg6#[K"@WjhZ}r[Sc
 un&sBCS,T[/3]KAeEnQ7lU)3Pn,0)G/6N
 -wyzj/MTd#A;r*(ds./df3r8Odaf?/<#r Parser Syntax Error Interpreter #
39. 39 Leveraging Instrumentation Based Feedback def parse_num(s,i): n = '' while s[i:] and is_digit(s[i]): n += s[i] i = i +1 return i,n def parse_paren(s, i): assert s[i] == '(' i, v = parse_expr(s, i+1) if s[i:] == '': raise Exception(s, i) assert s[i] == ')' return i+1, v def parse_expr(s, i = 0): expr = [] is_op = True while s[i:]: c = s[i] if c in string.digits: if not is_op: raise Exception(s,i) i,num = parse_num(s,i) expr.append(num) is_op = False elif c in ['+', '-', '*', '/']: if is_op: raise Exception(s,i) expr.append(c) is_op = True i = i + 1 elif c == '(': if not is_op: raise Exception(s,i) i, cexpr = parse_paren(s, i) expr.append(cexpr) is_op = False elif c == ')': break else: raise Exception(s,i) if is_op: raise Exception(s,i) return i, expr Parser Syntax Error Interpreter #
40. http://localhost:8888/notebooks/x1_1_TrackingAccess.ipynb 40
41. 41 Tracker for Access http://localhost:8888/notebooks/x1_1_TrackAccess.ipynb#xtstr
42. 42 AST Rewriting http://localhost:8888/notebooks/x1_1_TrackAccess.ipynb#InRewriter
43. 43 Find The Last Compared Index
44. 44 Speci fi cation Free Generator
45. 45 Speci fi cation Free Generator A ( 2 - B 9 ) 4 ) A ∉ (,+,-,1,2,3,4,5,6,7,8,9,0 B ∉ +,-,1,2,3,4,5,6,7,8,9,0,) ) ∉ +,-,1,2,3,4,5,6,7,8,9,0 (2-94)
47. 47 Limitation: Lack of control
48. 48 Constraining the Input Domain with Grammars
49. 49 Grammar
50. 50 Formal Languages Formal Language Descriptions 3. Regular Context Free Recursively Enumerable (Chomsky,1956) Easy to produce and parse Argument Stack Return Stack Easier to reason with
51. 51 Grammar <start> := <expr> <expr> := <expr> '+' <expr> | <expr> '-' <expr> | <expr> '/' <expr> | <expr> '*' <expr> | '(' <expr> ')' | <number> <number> := <integer> | <integer> '.' <integer> <integer>:= <digit> <integer> | <digit> <digit> := [0-9] Arithmetic expression grammar De f inition for <expr> <expr> key
52. 52 <start> := <expr> <expr> := <expr> '+' <expr> | <expr> '-' <expr> | <expr> '/' <expr> | <expr> '*' <expr> | '(' <expr> ')' | <number> <number> := <integer> | <integer> '.' <integer> <integer>:= <digit> <integer> | <digit> <digit> := [0-9] Grammar Arithmetic expression grammar Expansion Rule Terminal Symbol Nonterminal Symbol
53. http://localhost:8888/notebooks/x0_1_Grammars.ipynb 53
54. 54 Grammars For Parsing (8 / 3) * 49 <start> := <expr> <expr> := <expr> '+' <expr> | <expr> '-' <expr> | <expr> '/' <expr> | <expr> '*' <expr> | '(' <expr> ')' | <number> <number> := <integer> | <integer> '.' <integer> <integer>:= <digit> <integer> | <digit> <digit> := [0-9]
55. http://localhost:8888/notebooks/x0_3_Parser.ipynb 55
56. 56 Parsing is Surprisingly Integral to Fuzzing 1. Parsing is necessary for decomposing and recomposing inputs without affecting the validity of inputs constructed (8 / 3) * 49
57. 57 Parsing is Surprisingly Integral to Fuzzing 1. Parsing is necessary for decomposing and recomposing inputs without affecting the validity of inputs constructed (8 / 3) * 49 http://localhost:8888/notebooks/x0_3_Parser.ipynb#Arborist
58. 58 Parsing is Surprisingly Integral to Fuzzing 1. Parsing is necessary for decomposing and recomposing inputs without affecting the validity of inputs constructed (2 + 1) * 49 1 2 + (8 / 3) * 49
59. 59 Parsing is Surprisingly Integral to Fuzzing 2. Use parsers to mine the input distribution, and generate uncommon inputs
60. 60 Parsing is Surprisingly Integral to Fuzzing 2. Use parsers to extract the complexity of an input or a set of inputs 18439249 (8/3)*49 https://rahul.gopinath.org/post/2024/03/23/k-paths-for-context-free-grammars/
62. 62 Grammars 8.2 - 27 - -9 / +((+9 * --2 + --+-+- ((-1 * +(8 - 5 - 6)) * (-(a-+(((+(4) )))) - ++4) / +(-+---((5.6 - --(3 * -1.8 * +(6 * +-(((-(-6) * ---+6)) / +--(+-+-7 * (-0 * (+(((((2)) + 8 - 3 - ++9.0 + ---(--+7 / (1 / +++6.37) + (1) / 482) / +++-+0)))) + 8.2 - 27 - -9 / +((+9 * --2 + --+-+-((-1 * + (8 - 5 - 6)) * (-(a-+(((+(4))))) - + +4) / +(-+---((5.6 - --(3 * -1.8 * + (6 * +-(((-(-6) * ---+6)) / +--(+-+- 7 * (-0 * (+(((((2)) + 8 - 3 - ++9.0 + ---(--+7 / (1 / +++6.37) + (1) / 482) / +++-+0)))) * -+5 + 7.513)))) - (+1 / ++((-84)))))))) * ++5 / +-(- -2 - -++-9.0)))) / 5 * --++090 + * - +5 + 7.513)))) - (+1 / ++((-84)))))) )) * 8.2 - 27 - -9 / +((+9 * --2 + - -+-+-((-1 * +(8 - 5 - 6)) * (-(a-+(( (+(4))))) - ++4) / +(-+---((5.6 - -- (3 * -1.8 * +(6 * +-(((-(-6) * ---+6 )) / +--(+-+-7 * (-0 * (+(((((2)) + 8 - 3 - ++9.0 + ---(--+7 / (1 / +++6 .37) + (1) / 482) / +++-+0)))) * -+5 + 7.513)))) - (+1 / ++((-84)))))))) * ++5 / +-(--2 - -++-9.0)))) / 5 * --++090 ++5 / +-(--2 - -++-9.0)))) / 5 * --++090 <start> := <expr> <expr> := <expr> '+' <expr> | <expr> '-' <expr> | <expr> '/' <expr> | <expr> '*' <expr> | '(' <expr> ')' | <number> <number> := <integer> | <integer> '.' <integer> <integer>:= <digit> <integer> | <digit> <digit> := [0-9] For Fuzzing (Hanford 1970) (Purdom 1972)
63. 63 Grammars As effective producers Interpreter Parser ✘ ✔ 8.2 - 27 - -9 / +((+9 * --2 + --+-+- ((-1 * +(8 - 5 - 6)) * (-(a-+(((+(4) )))) - ++4) / +(-+---((5.6 - --(3 * -1.8 * +(6 * +-(((-(-6) * ---+6)) / +--(+-+-7 * (-0 * (+(((((2)) + 8 - 3 - ++9.0 + ---(--+7 / (1 / +++6.37) + (1) / 482) / +++-+0)))) + 8.2 - 27 - -9 / +((+9 * --2 + --+-+-((-1 * + (8 - 5 - 6)) * (-(a-+(((+(4))))) - + +4) / +(-+---((5.6 - --(3 * -1.8 * + (6 * +-(((-(-6) * ---+6)) / +--(+-+- 7 * (-0 * (+(((((2)) + 8 - 3 - ++9.0 + ---(--+7 / (1 / +++6.37) + (1) / 482) / +++-+0)))) * -+5 + 7.513)))) - (+1 / ++((-84)))))))) * ++5 / +-(- -2 - -++-9.0)))) / 5 * --++090 + * - +5 + 7.513)))) - (+1 / ++((-84)))))) )) * 8.2 - 27 - -9 / +((+9 * --2 + - -+-+-((-1 * +(8 - 5 - 6)) * (-(a-+(( (+(4))))) - ++4) / +(-+---((5.6 - -- (3 * -1.8 * +(6 * +-(((-(-6) * ---+6 )) / +--(+-+-7 * (-0 * (+(((((2)) + 8 - 3 - ++9.0 + ---(--+7 / (1 / +++6 .37) + (1) / 482) / +++-+0)))) * -+5 + 7.513)))) - (+1 / ++((-84)))))))) * ++5 / +-(--2 - -++-9.0)))) / 5 * --++090 ++5 / +-(--2 - -++-9.0)))) / 5 * --++090
64. 64 Grammars <start> := <expr> <expr> := <expr> '+' <expr> | <expr> '-' <expr> | <expr> '/' <expr> | <expr> '*' <expr> | '(' <expr> ')' | <number> <number> := <integer> | <integer> '.' <integer> <integer>:= <digit> <integer> | <digit> <digit> := [0-9] As efficient producers def start(): expr() def expr(): match (random() % 6): case 0: expr(); print('+'); expr() case 1: expr(); print('-'); expr() case 2: expr(); print('/'); expr() case 3: expr(); print('*'); expr() case 4: print('('); expr(); print(')') case 5: number() def number(): match (random() % 2): case 0: integer() case 1: integer(); print('.'); integer() def integer(): match (random() % 2): case 0: digit(); integer() case 1: digit() def digit(): match (random() % 10): case 0: print('0') case 1: print('1') case 2: print('2') case 3: print('3') case 4: print('4') case 5: print('5') case 6: print('6') case 7: print('7') Compiled Grammar (F1)
65. http://localhost:8888/notebooks/x0_2_GrammarFuzzer.ipynb 65
66. 66 Where to Get the Input Grammar From?
67. 67 https://www.fuzzingbook.org/html/GrammarMiner.html AUTOGRAM Where to Get the Grammar From? Hand-written parsers already encode the grammar
68. 68 Where to Get the Grammar From? Handwritten parsers contain the parse structure key value key value scheme parse_scheme parse_hostpath parse_querystring parse_fragment domain TLD subdomain parse_host subdirectory parse_fslocation binary parse_binaryname parameters parse_parameters parse_url
69. 69 Where to Get the Grammar From? Mining Grammar from a hand-written parser https://www.example.com/forum/questions/cgi?tag=networking&order=newwest#top key value key value split scheme parse_scheme host path parse_hostpath query string parse_querystring fragment parse_fragment domain TLD subdomain parse_host subdirectory parse_fslocation binary parse_binaryname parameters parse_parameters With Dynamic Data Flow Analysis parseurl
70. 70 http://user:pass@www.google.com:80/?q=path#ref urlparse:url = 'http://user:pass@www.google.com:80/?q=path#ref' urlsplit:scheme = 'http' urlsplit:netloc = 'user:pass@www.google.com:80' urlsplit:fragment = 'ref' urlsplit:query = 'q=path' https://soft-eng.sydney.edu.au:80/ urlparse:url = 'https://soft-eng.sydney.edu.au:80/' urlsplit:scheme = 'https' urlsplit:netloc = 'soft-eng.sydney.edu.au:80' http://www.fuzzingbook.org/#News urlparse:url = 'http://www.fuzzingbook.org/#News' urlsplit:scheme = 'http' urlsplit:netloc = 'www.fuzzingbook.org' urlsplit:fragment = 'News' Mining with Dynamic Data Flow Analysis
71. 71 { '<urlparse:url>': [ ['<urlsplit:scheme>', '://', '<urlsplit:netloc>', '/?', '<urlsplit:query>', '#', '<urlsplit:fragment>'], ['<urlsplit:scheme>', '://', '<urlsplit:netloc>', '/#','<urlsplit:fragment>'], ['<urlsplit:scheme>', '://', '<urlsplit:netloc>', '/']], '<urlsplit:scheme>' : [ ['http'], ['http']], '<urlsplit:netloc>': [ ['user:pass@www.google.com:80'], ['www.fuzzingbook.org'], ['soft-eng.sydney.edu.au']], '<urlsplit:query>' : [ ['q=path']], '<urlsplit:fragment>' : [ ['ref'], ['News']],
 } http://user:pass@www.google.com:80/?q=path#ref urlparse:url = 'http://user:pass@www.google.com:80/?q=path#ref' urlsplit:scheme = 'http' urlsplit:netloc = 'user:pass@www.google.com:80' urlsplit:fragment = 'ref' urlsplit:query = 'q=path' https://soft-eng.sydney.edu.au:80/ urlparse:url = 'https://soft-eng.sydney.edu.au:80/' urlsplit:scheme = 'https' urlsplit:netloc = 'soft-eng.sydney.edu.au:80' http://www.fuzzingbook.org/#News urlparse:url = 'http://www.fuzzingbook.org/#News' urlsplit:scheme = 'http' urlsplit:netloc = 'www.fuzzingbook.org' urlsplit:fragment = 'News' Mining with Dynamic Data Flow Analysis
72. { '<urlparse:url>': [ ['<urlsplit:scheme>', '://', '<urlsplit:netloc>', '/?', '<urlsplit:query>', '#', '<urlsplit:fragment>'], ['<urlsplit:scheme>', '://', '<urlsplit:netloc>', '/#','<urlsplit:fragment>'], ['<urlsplit:scheme>', '://', '<urlsplit:netloc>', '/']], '<urlsplit:scheme>' : [ ['http'], ['http']], '<urlsplit:netloc>': [ ['user:pass@www.google.com:80'], ['www.fuzzingbook.org'], ['soft-eng.sydney.edu.au']], '<urlsplit:query>' : [ ['q=path']], '<urlsplit:fragment>' : [ ['ref'], ['News']],
 } Limitations • Poor accuracy in most handwritten parsers • Handwritten parsers are not often well formed • Control flow is ignored Mining with Dynamic Data Flow Analysis
73. http://localhost:8888/notebooks/x2_0_MiningGrammar.ipynb 73
74. 74 MIMID Where to Get the Grammar From? Hand-written parsers already encode the grammar http://localhost:8888/notebooks/x2_0_MiningGrammar.ipynb
75. 75 Where to Get the Grammar From? 1. Extract the input string accesses 2. Attach control fl ow information Hand-written parsers already encode the grammar Each control- fl ow structure gets wrapped in a context-manager -If conditionals -Loops -Subroutines
76. 76 Instrumentation by AST rewrite http://localhost:8888/notebooks/x2_0_MiningGrammar.ipynb#Context-Mangers
77. 77 Track Where Each Character Index was Accessed http://localhost:8888/notebooks/x2_0_MiningGrammar.ipynb#Rewriting-the-source-to-track-control- fl ow-and-taints.
78. 78 How to Extract This Grammar? • Inputs + control fl ow -> Dynamic Control Dependence Trees • DCD Trees -> Parse Tree http://localhost:8888/notebooks/x2_0_MiningGrammar.ipynb#Mining-the-Traces-Generated
79. 79 Control Dependence Graph Statement B is control dependent on A if A determines whether B executes. def parse_csv(s,i): while s[i:]: if is_digit(s[i]): n,j = num(s[i:]) i = i+j else: comma(s[i]) i += 1 CDG for parse_csv while: determines whether if: executes
80. 80 def parse_csv(s,i): while s[i:]: if is_digit(s[i]): n,j = num(s[i:]) i = i+j else: comma(s[i]) i += 1 CDG for parse_csv Dynamic Control Dependence Tree Each statement execution is represented as a separate node DCD Tree for call parse_csv()
81. 81 def parse_csv(s,i): while s[i:]: if is_digit(s[i]): n,j = num(s[i:]) i = i+j else: comma(s[i]) i += 1 '1' '2' ',' DCD Tree ~ Parse Tree •No tracking beyond input bu ff er •Characters are attached to nodes where they are accessed last "12," "12,"
82. 82 •Characters are attached to nodes where they are accessed last http://localhost:8888/notebooks/x2_0_MiningGrammar.ipynb#reparsing-behaviour
83. 83 http://localhost:8888/notebooks/x2_0_MiningGrammar.ipynb#The-Complete-Miner The Complete Parse Tree Miner
84. 84 def is_digit(i): return i in '0123456789' def parse_num(s,i): n = '' while s[i:] and is_digit(s[i]): n += s[i] i = i +1 return i,n def parse_paren(s, i): assert s[i] == '(' i, v = parse_expr(s, i+1) if s[i:] == '': raise Ex(s, i) assert s[i] == ')' return i+1, v def parse_expr(s, i = 0): expr, is_op = [], True while s[i:]: c = s[i] if isdigit(c): if not is_op: raise Ex(s,i) i,num = parse_num(s,i) expr.append(num) is_op = False elif c in ['+', '-', '*', '/']: if is_op: raise Ex(s,i) expr.append(c) is_op, i = True, i + 1 elif c == '(': if not is_op: raise Ex(s,i) i, cexpr = parse_paren(s, i) expr.append(cexpr) is_op = False elif c == ')': break else: raise Ex(s,i) if is_op: raise Ex(s,i) return i, expr 9+3/4 Parse tree for parse_expr('9+3/4')
85. 85 http://localhost:8888/notebooks/x2_0_MiningGrammar.ipynb#Generalize-Nodes
86. 86 9+3/4 Identifying Compatible Nodes Which nodes correspond to the same nonterminal http://localhost:8888/notebooks/x2_0_MiningGrammar.ipynb#Generalize-Loops
87. 87 (9 + 1) * 3 3 * (9 + 1)
88. 88 9 + 1 3 * (9 + 1)
89. 89 3 (9 + 1) * 3 * (9 + 1)
90. 90 3*(1) 1
91. 91 3*(1) 1 <parse_expr> := <while 1:1> <while 1:0> <while 1:1> <while 1:1> <parse_expr> :=
92. <parse_expr> := <while 1:1> <while 1:0> <while 1:1> | <while 1:1> <while 1:0> <while 1:1> <while 1:0> <while 1:1> <while 1:0> <while 1:1> | <while 1:1> <while 1:0> <while 1:1> <while 1:0> <while 1:1> | <while 1:1> <while 1:1> := <if 1:1> <if 1:1> := <parse_num> | <parse_paren> <parse_num> := <is_digit> <is_digit> := '3' | '1' <parse_paren>:= '(' <parse_expr> ')' <while 1:0> := <if 1:0> <if 1:0> := '*' 92
93. <parse_expr> := <while_s> <while_s> := <while_1:1> <while_1:0> <while_s> | <while_1:1> <parse_expr> := <while 1:1> <while 1:0> <while 1:1> | <while 1:1> <while 1:0> <while 1:1> <while 1:0> <while 1:1> <while 1:0> <while 1:1> | <while 1:1> <while 1:0> <while 1:1> <while 1:0> <while 1:1> | <while 1:1> <while 1:1> := <if 1:1> <if 1:1> := <parse_num> | <parse_paren> <parse_num> := <is_digit> <is_digit> := '3' | '1' <parse_paren>:= '(' <parse_expr> ')' <while 1:0> := <if 1:0> <if 1:0> := '*' 93 Generalizing Loops with Regular Inference
94. 94 http://localhost:8888/notebooks/x2_0_MiningGrammar.ipynb#Generating-a-Grammar
95. 95 def is_digit(i): return i in '0123456789' def parse_num(s,i): n = '' while s[i:] and is_digit(s[i]): n += s[i] i = i +1 return i,n def parse_paren(s, i): assert s[i] == '(' i, v = parse_expr(s, i+1) if s[i:] == '': raise Ex(s, i) assert s[i] == ')' return i+1, v def parse_expr(s, i = 0): expr, is_op = [], True while s[i:]: c = s[i] if isdigit(c): if not is_op: raise Ex(s,i) i,num = parse_num(s,i) expr.append(num) is_op = False elif c in ['+', '-', '*', '/']: if is_op: raise Ex(s,i) expr.append(c) is_op, i = True, i + 1 elif c == '(': if not is_op: raise Ex(s,i) i, cexpr = parse_paren(s, i) expr.append(cexpr) is_op = False elif c == ')': break else: raise Ex(s,i) if is_op: raise Ex(s,i) return i, expr <START> := <parse_expr.0-0-c> <parse_expr.0-0-c> := <parse_expr.0-1-s><parse_expr.0> | <parse_expr.0> <parse_expr.0-1-s> := <parse_expr.0><parse_expr.0-2> | <parse_expr.0><parse_expr.0-2><parse_expr.0-1-s> <parse_expr.0> := '(' <parse_expr.0-0-c> ')' | <parse_num.0-1-s> <parse_expr.0-2> := '*' | '+' | '-' | '/' <parse_num.0-1-s> := <is_digit.0-0-c> | <is_digit.0-0-c><parse_num.0-1-s> <is_digit.0-0-c> : [0-9] calc.py Recovered Arithmetic Grammar
96. 96 8.2 - 27 - -9 / +((+9 * --2 + --+-+- ((-1 * +(8 - 5 - 6)) * (-(a-+(((+(4) )))) - ++4) / +(-+---((5.6 - --(3 * -1.8 * +(6 * +-(((-(-6) * ---+6)) / +--(+-+-7 * (-0 * (+(((((2)) + 8 - 3 - ++9.0 + ---(--+7 / (1 / +++6.37) + (1) / 482) / +++-+0)))) + 8.2 - 27 - -9 / +((+9 * --2 + --+-+-((-1 * + (8 - 5 - 6)) * (-(a-+(((+(4))))) - + +4) / +(-+---((5.6 - --(3 * -1.8 * + (6 * +-(((-(-6) * ---+6)) / +--(+-+- 7 * (-0 * (+(((((2)) + 8 - 3 - ++9.0 + ---(--+7 / (1 / +++6.37) + (1) / 482) / +++-+0)))) * -+5 + 7.513)))) - (+1 / ++((-84)))))))) * ++5 / +-(- -2 - -++-9.0)))) / 5 * --++090 + * - +5 + 7.513)))) - (+1 / ++((-84)))))) )) * 8.2 - 27 - -9 / +((+9 * --2 + - -+-+-((-1 * +(8 - 5 - 6)) * (-(a-+(( (+(4))))) - ++4) / +(-+---((5.6 - -- (3 * -1.8 * +(6 * +-(((-(-6) * ---+6 )) / +--(+-+-7 * (-0 * (+(((((2)) + 8 - 3 - ++9.0 + ---(--+7 / (1 / +++6 .37) + (1) / 482) / +++-+0)))) * -+5 + 7.513)))) - (+1 / ++((-84)))))))) * ++5 / +-(--2 - -++-9.0)))) / 5 * --++090 ++5 / +-(--2 - -++-9.0)))) / 5 * --++090 <START> := <parse_expr.0-0-c> <parse_expr.0-0-c> := <parse_expr.0-1-s><parse_expr.0> | <parse_expr.0> <parse_expr.0-1-s> := <parse_expr.0><parse_expr.0-2> | <parse_expr.0><parse_expr.0-2><parse_expr.0-1-s> <parse_expr.0> := '(' <parse_expr.0-0-c> ')' | <parse_num.0-1-s> <parse_expr.0-2> := '*' | '+' | '-' | '/' <parse_num.0-1-s> := <is_digit.0-0-c> | <is_digit.0-0-c><parse_num.0-1-s> <is_digit.0-0-c> : [0-9]
97. 97 <START> ::= <json_raw> <json_raw> ::= '"' <json_string'> | '[' <json_list'> | '{' <json_dict'> | <json_number'> | 'true' | 'false' | 'null' <json_number'> ::= <json_number>+ | <json_number>+ 'e' <json_number>+ <json_number> ::= '+' | '-' | '.' | [0-9] | 'E' | 'e' <json_string'> ::= <json_string>* '"' <json_list'> ::= ']' | <json_raw> (','<json_raw>)* ']' | ( ',' <json_raw>)+ (',' <json_raw>)* ']' <json_dict'> ::= '}' | ( '"' <json_string'> ':' <json_raw> ',' )* '"'<json_string'> ':' <json_raw> '}' <json_string> ::= ' ' | '!' | '#' | '$' | '%' | '&' | ''' | '*' | '+' | '-' | ',' | '.' | '/' | ':' | ';' | '<' | '=' | '>' | '?' | '@' | '[' | ']' | '^' | '_', ''',| '{' | '|' | '}' | '~' | '[A-Za-z0-9]' | '\' <decode_escape> <decode_escape> ::= '"' | '/' | 'b' | 'f' | 'n' | 'r' | 't' stm.next() if expect_key: raise JSONError(E_DKEY, stm, stm.pos) if c == '}': return result expect_key = 1 continue # parse out a key/value pair elif c == '"': key = _from_json_string(stm) stm.skipspaces() c = stm.next() if c != ':': raise JSONError(E_COLON, stm, stm.pos) stm.skipspaces() val = _from_json_raw(stm) result[key] = val expect_key = 0 continue raise JSONError(E_MALF, stm, stm.pos) def _from_json_raw(stm): while True: stm.skipspaces() c = stm.peek() if c == '"': return _from_json_string(stm) elif c == '{': return _from_json_dict(stm) elif c == '[': return _from_json_list(stm) elif c == 't': return _from_json_fixed(stm, 'true', True, E_BOOL) elif c == 'f': return _from_json_fixed(stm, 'false', False, E_BOOL) elif c == 'n': return _from_json_fixed(stm, 'null', None, E_NULL) elif c in NUMSTART: return _from_json_number(stm) raise JSONError(E_MALF, stm, stm.pos) def from_json(data): stm = JSONStream(data) return _from_json_raw(stm) microjson.py Recovered JSON grammar
101. 101 The Oracle Problem
102. 102 The Oracle Problem • You fuzz the program with an input. • The program runs. • It fi nishes. Was the answer correct?
103. 103 The Oracle Problem Was the answer correct? Option 1: Di ff erential Testing Verify input/output against other implementations
104. 104 The Oracle Problem Was the answer correct? Option 1: Di ff erential Testing Constraint: Need access to di ff erent program implementations (This is the basis for regression testing)
105. 105 The Oracle Problem Was the answer correct? Option 1: Di ff erential Testing (This is the basis for regression testing) Option 2: Property Based Testing Constraint: Need access to di ff erent program implementations assert f(x) == y
106. 106 The Oracle Problem Was the answer correct? Option 1: Di ff erential Testing (This is the basis for regression testing) Option 2: Property Based Testing 3: Metamorphic Testing Constraint: Need access to di ff erent program implementations assert f(x + c) == f(x) + c
107. 107 The Oracle Problem Was the answer correct? Option 1: Di ff erential Testing (This is the basis for regression testing) Option 2: Property Based Testing 3: Metamorphic Testing Constraint: Need access to di ff erent program implementations Constraint: Need speci fi cations
 (or intelligence)
108. 108 The Oracle Problem Was the answer correct? Option 1: Di ff erential Testing (This is the basis for regression testing) Option 2: Property Based Testing 3: Metamorphic Testing Constraint: Need access to di ff erent program implementations Constraint: Need speci fi cations
 (or intelligence) Option 4: Symbolic Execution
109. 109 The Oracle Problem Was the answer correct? Option 1: Di ff erential Testing (This is the basis for regression testing) Option 2: Property Based Testing 3: Metamorphic Testing Constraint: Need access to di ff erent program implementations Constraint: Need speci fi cations
 (or intelligence) Option 4: Symbolic Execution
 5: Fuzzing
110. 110 The Oracle Problem Was the answer correct? Option 1: Di ff erential Testing (This is the basis for regression testing) Option 2: Property Based Testing 3: Metamorphic Testing Constraint: Need access to di ff erent program implementations Constraint: Need speci fi cations
 (or intelligence) Option 4: Symbolic Execution
 5: Fuzzing Constraint: Limited general oracles
 (e.g. sanitizers)
111. 111 The Oracle Problem: Solutions
112. 112 The Oracle Problem: Solutions Normal execution establishes patterns: variables stay within ranges, relationships hold.
 A rare execution is one where those patterns break. That breakage is a potential bug.
113. Option 1: Di ff erential Testing Option 2: Property Based Testing 3: Metamorphic Testing Constraint: Need access to di ff erent program implementations Constraint: Need speci fi cations
 (or intelligence) Option 4: Symbolic Execution
 5: Fuzzing Constraint: Limited general oracles
 (e.g. sanitizers) 113 The Oracle Problem Was the answer correct? (This is the basis for regression testing) Option 6: Approximate Oracles Daikon Runtime Invariant Miner Ernst et al. 2001
114. 114 A Runtime Invariant Miner http://localhost:8888/notebooks/RoadMap.ipynb#x6_0_InvariantMining
115. 115 Daikon: Core Idea def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene • Run the program on a bunch of inputs • Watch every variable at key points. • Find properties that always hold. triangle(3, 3, 3) → ENTER: a=3, b=3, c=3 EXIT: return Equilateral triangle(3, 3, 4) → ENTER: a=3, b=3, c=4 EXIT: return Isosceles triangle(2, 2, 4) → ENTER: a=3, b=3, c=4 EXIT: return Isosceles
116. 116 Daikon: Core Idea def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene • Run the program on a bunch of inputs • Watch every variable at key points. • Find properties that always hold. triangle(3, 3, 3) → ENTER: a=3, b=3, c=3 EXIT: return Equilateral triangle(3, 3, 4) → ENTER: a=3, b=3, c=4 EXIT: return Isosceles triangle(2, 2, 4) → ENTER: a=2, b=2, c=4 EXIT: return Isosceles After observing many runs: a >= 0 ✓ held every time → likely invariant a == b ✓ held every time → likely invariant
117. 117 Daikon: Core Idea def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene • Run the program on a bunch of inputs • Watch every variable at key points. • Find properties that always hold. triangle(3, 3, 3) → ENTER: a=3, b=3, c=3 EXIT: return Equilateral triangle(3, 3, 4) → ENTER: a=3, b=3, c=4 EXIT: return Isosceles triangle(2, 2, 3) → ENTER: a=2, b=2, c=3 EXIT: return Isosceles After observing many runs: a >= 0 ✓ held every time → likely invariant a == b ✘ held every time → likely invariant a + b < c ✓ held every time → likely invariant
118. 118 Daikon: Core Idea def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene • Run the program on a bunch of inputs • Watch every variable at key points. • Find properties that always hold. triangle(3, 3, 3) → ENTER: a=3, b=3, c=3 EXIT: return Equilateral triangle(3, 3, 4) → ENTER: a=3, b=3, c=4 EXIT: return Isosceles triangle(2, 2, 3) → ENTER: a=2, b=2, c=3 EXIT: return Isosceles triangle(2, 3, 4) → ENTER: a=2, b=3, c=4 EXIT: return Scalene After observing many runs: a >= 0 ✓ held every time → likely invariant a == b ✘ not held → discarded a + b < c ✓ held every time → likely invariant
119. 119 Daikon: Core Idea def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene • Run the program on a bunch of inputs • Watch every variable at key points. • Find properties that always hold. triangle(3, 3, 3) → ENTER: a=3, b=3, c=3 EXIT: return Equilateral triangle(3, 3, 4) → ENTER: a=3, b=3, c=4 EXIT: return Isosceles triangle(2, 2, 3) → ENTER: a=2, b=2, c=3 EXIT: return Isosceles triangle(2, 3, 4) → ENTER: a=2, b=3, c=4 EXIT: return Scalene After observing many runs: a >= 0 ✓ held every time → likely invariant a + b < c ✓ held every time → likely invariant
120. 120 Daikon: Core Idea def triangle(a, b, c): if a == b: if b == c: return Equilateral else: return Isosceles else: if b == c: return Isosceles else: if a == c: return Isosceles else: return Scalene •These are likely invariants, not proofs.
 •They describe what the code does, not what it should do.
121. 121 Program Points • variable values (and invariants) are anchored to where they were observed.
 
 triangle:::ENTER ← variable values when function is called
 triangle:::EXIT ← variable values (+ return value) when it returns • The same property at `ENTER` and at `EXIT` are two separate invariants. • A variable might be `>= 0` on the way in but not on the way out. • This is Daikon's key design decision: invariants are always located. http://localhost:8888/notebooks/x6_0_SimpleInvariantMiner.ipynb#Program-Points
122. 122 Candidate Templates The miner generates templates for every variable it sees, then kills them when they fail. Unary (one variable): - x >= 0 - type(x) is int - x is not None Binary (pairs of variables): - x == y
 - x <= y
 - x >= y 
 Ternary and so on. Any template that fails even once is discarded. Survivors are your invariants. The template set is the limit of what the miner can discover. Add templates, discover more. http://localhost:8888/notebooks/x6_0_SimpleInvariantMiner.ipynb#Candidate-Invariants
123. 123 Suppression: Cutting the Noise The suppression step removes weaker invariants dominated by stronger ones: x == y → 
 x <= y 
 x >= y
 x is not None
 y is not None http://localhost:8888/notebooks/x6_0_SimpleInvariantMiner.ipynb#Suppression
124. What To Do with Invariants Invariants become assertions in your test suite def test_triangle_invariants(a, b, c): result = triangle(a, b, c) assert a >= 0 assert b >= 0 assert c >= 0 assert isinstance(result, str) Mined from triangle runs — auto-generated oracle
125. 125 Leveraging Uncommon Executions An invariant violation means the program has entered a state it has never been in before. • Rare executions are worth investigating --- because their correctness has never been veri fi ed. • If unusual, fl ag the current input as a seed for further fuzzing • If extreme, fl ag for manual inspection
127. 127 Now for something completely different.
128. 128 Variety of Fuzzers Publications from 2023-2024
129. 129 Which Fuzzer Should We Use? Exponential growth in fuzzing literature Cumulative publications in fuzzing Cumulative published vulnerabilities
130. 130 Which Evaluation Metric Should I Use? • Comparable • Ground truth • Unbiased • Budget friendly
131. 131 New CVEs Found? CVE, short for Common Vulnerabilities and Exposures, is a list of publicly disclosed computer security fl aws.
132. 132 • Comparable • Ground truth • Unbiased • Budget friendly Which Fuzzer Should I Use? New CVEs ? ?
133. 133 Evaluate Fuzzers Based on the Speci fi ed Goal Coverage?
134. 134 • Gets saturated quickly (Signal is lost) • Is reachability sufficient? • Limited evaluation of complex input conditions involved in expressing bugs • Are more complex coverage techniques needed? Coverage?
135. 135 • Comparable • Ground truth • Unbiased • Budget friendly Which Fuzzer Should I Use? New CVEs ? ? Coverage
136. 136 Synthetic Bugs? Evaluate Fuzzers Based on the Speci fi ed Goal
137. 137 Synthetic Bugs? LAVA: Large-scale Automated Vulnerability Addition • What kind of bugs should be seeded? • What about unknown kinds of bugs? • Where should these bugs be seeded?
138. 138 • Comparable • Ground truth • Unbiased • Budget friendly Which Fuzzer Should I Use? New CVEs ? ? Coverage Synthetic (LAVA) ? ?
139. 139 Curated Bug Benchmarks? Evaluate Fuzzers Based on the Speci fi ed Goal
140. 140 Known Bug Benchmarks? E.g. Magma
141. 141 Known Bug Benchmarks? • Gets outdated quickly (overfitting by fuzzers) • Significant effort in creating and maintaining • Biased as to where and what kind of bugs are present "When a measure becomes a target it ceases to be a good measure" Goodhart's law
142. 142 • Comparable • Ground truth • Unbiased • Budget friendly Which Fuzzer Should I Use? New CVEs ? ? Coverage Synthetic (LAVA) ? ? Benchm arks (M AGM A) ?
143. 143 New Bugs in Existing Benchmarks Evaluate Fuzzers Based on the Speci fi ed Goal
144. 144 New Bugs in Existing Benchmarks • No ground truth • Bug distribution is dependent on external factors • Can lead researchers to postpone publication of vulnerabilities • Feedback can't be used to decide budgeting.
145. 145 • Comparable • Ground truth • Unbiased • Budget friendly Which Fuzzer Should I Use? New CVEs ? ? Coverage Synthetic (LAVA) ? ? Benchm arks (M AGM A) ? New Bugs in old Benchm arks ? Can
147. 147 Fuzzing Your Fuzzer a.k.a. Mutation Testing
148. 148 unsigned int len = message_length(msg); if (len < MAX_BUF_LEN) { copy_message(msg); } else { // Invalid length, handle error } What Is Mutation Testing? IDEA: Induce a program variation with each valid token replacement
149. 149 unsigned int len = message_length(msg); if (len < MAX_BUF_LEN + 1) { copy_message(msg); } else { // Invalid length, handle error } What Is Mutation Testing? Fixes for independent bugs are almost always simple. Finite syntactic size for faults
 (aka. competent programmer hypothesis): Gopinath, Jensen, and Groce Mutations: How Close are they to Real Faults? 2014 ISSRE
150. 150 unsigned int len = message_length(msg); if (len < MAX_BUF_LEN) { copy_message(msg); } else { // Invalid length, handle error } What Is Mutation Testing? (1) IDEA: Induce a program variation with each valid token replacement
151. 151 unsigned int len = message_length(msg); if (len < >= MAX_BUF_LEN) { copy_message(msg); } else { // Invalid length, handle error } What Is Mutation Testing? (2) IDEA: Induce a program variation with each valid token replacement
152. 152 unsigned int len = message_length(msg); if (len < MAX_BUF_LEN + 1) { copy_message(msg); } else { // Invalid length, handle error } What Is Mutation Testing? (3) IDEA: Induce a program variation with each valid token replacement
153. 153 unsigned int len = message_length(msg); if (len < MAX_BUF_LEN + 16) { copy_message(msg); } else { // Invalid length, handle error } What Is Mutation Testing? Complex bugs are almost always coupled to simpler bugs. Finite semantic depth for failures
 aka. Coupling e ff ect hypothesis: Gopinath, Jensen, and Groce Mutations: How Close are they to Real Faults? 2017 ICST
155. 155 Mutation Testing Process m4 m3 m2 m1 I1 I2 I4 I3 I5 Generate Mutants Generate Fuzz Inputs > Detect Differences from Original > Mutation Score = Detected Mutants / Generated Mutants Original
156. 156 M mutants (1 input) Number of Mutants Executions Total Campaign Effort for mutation testing = MxN program executions Mutation Testing Challenge N inputs (1 mutant) Number of Inputs Executions
157. 157 What Is The Problem? Computation: Fuzzing -- More executions the better Mutation testing -- Execute each input per mutation Solution: 1) Perform coverage analysis fi rst; remove trivial mutants 2) Evaluate independent (non-interacting) mutations simultaneously 3) Remove redundancy in executions
158. https://rahul.gopinath.org Traditional execution 158 m1 m2 m3 Setup for T1 Setup for T2 Actual tests Mutants are executed in parallel But a majority of time spent in initialization (an average 7 times the test execution time) (Bell 2014)
159. https://rahul.gopinath.org Split-Stream Execution 159 Setup for T1 Setup for T2 Actual tests Execute tests in parallel Fork off mutants as they are encountered T1 T2 Gopinath, Jensen, Groce “Topsy Turvy: A faster and smarter algorithm” ICSE 2016
160. https://rahul.gopinath.org Split-Stream Execution 160
161. 161 • Comparable • Ground truth • Unbiased • Budget friendly Mutation Testing New CVEs ? ? Coverage Synthetic (LAVA) ? ? Benchm arks (M AGM A) ? New Bugs in old Benchm arks ? M utation Testing ?
163. 163 Seeded Fault Benchmarks • Easy to fine-tune a fuzzer to overfit • Faults are rarely similar to real faults • Based on bugs we know about! • Human bias in bug curation • Limited supply • Bug interactions requiring deduplication
164. 164 Mutation Analysis • Easy to fine-tune a fuzzer to overfit • Faults are rarely similar to real faults • Based on bugs we know about! • Human bias in bug curation • Limited supply • Bug interactions requiring deduplication
165. 165 Mutation Analysis • Easy to fine-tune a fuzzer to overfit Very large number of faults • Faults are rarely similar to real faults • Based on bugs we know about! • Human bias in bug curation • Limited supply • Bug interactions requiring deduplication
166. 166 Mutation Analysis • Not easy to fine-tune • Faults are rarely similar to real faults • Based on bugs we know about! • Human bias in bug curation • Limited supply • Bug interactions requiring deduplication
167. 167 Mutation Analysis • Not easy to fine-tune • Faults are rarely similar to real faults Evidence that mutants are similar to real faults. • Based on bugs we know about! • Human bias in bug curation • Limited supply • Bug interactions requiring deduplication
168. 168 Mutation Analysis • Not easy to fine-tune • Evidence that mutants are similar to real faults. • Based on bugs we know about! • Human bias in bug curation • Limited supply • Bug interactions requiring deduplication
169. 169 Mutation Analysis • Not easy to fine-tune • Evidence that mutants are similar to real faults. • Based on bugs we know about! All possible faults, including unknown ones! • Human bias in bug curation • Limited supply • Bug interactions requiring deduplication
170. 170 Mutation Analysis • Not easy to fine-tune • Evidence that mutants are similar to real faults. • All possible faults, including unknown ones! • Human bias in bug curation • Limited supply • Bug interactions requiring deduplication
171. 171 Mutation Analysis • Not easy to fine-tune • Evidence that mutants are similar to real faults. • All possible faults, including unknown ones! • Human bias in bug curation No human bias in introduced faults! • Limited supply • Bug interactions requiring deduplication
172. 172 Mutation Analysis • Not easy to fine-tune • Evidence that mutants are similar to real faults. • All possible faults, including unknown ones! • No human bias in introduced faults! • Limited supply • Bug interactions requiring deduplication
173. 173 Mutation Analysis • Not easy to fine-tune • Evidence that mutants are similar to real faults. • All possible faults, including unknown ones! • No human bias in introduced faults! • Limited supply As many as required! Including higher order ones! • Bug interactions requiring deduplication
174. 174 Mutation Analysis • Not easy to fine-tune • Evidence that mutants are similar to real faults. • All possible faults, including unknown ones! • No human bias in introduced faults! • As many as required! Including higher order ones! • Bug interactions requiring deduplication
175. 175 Mutation Analysis • Not easy to fine-tune • Evidence that mutants are similar to real faults. • All possible faults, including unknown ones! • No human bias in introduced faults! • As many as required! Including higher order ones! • Bug interactions requiring deduplication All mutants are evaluated independently
176. 176 Mutation Analysis • Not easy to fine-tune • Evidence that mutants are similar to real faults. • All possible faults, including unknown ones! • No human bias in introduced faults! • As many as required! Including higher order ones! • All mutants are evaluated indepdently