Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Sign up for free
Menu
Search
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Pricing
Search
Sign in
Sign up for free
SleeplyPickle.pdf
Search
CHOU,YUNSHIUAN
March 31, 2025
Research
37
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
SleeplyPickle.pdf
CHOU,YUNSHIUAN
March 31, 2025
Other Decks in Research
See All in Research
Anthropic が提案する LLM の内部状態を自然言語で説明可能にした Natural Language Autoencoders / Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations
shunk031
0
210
NLP colloquium: AI Safety Survey
kanekomasahiro
2
1.1k
Fukui Shibiten 39 - AI Art
butchi
0
190
SAKURAONE:An Open Ethernet-based AI HPC System And Its Observed Workload Dynamicsin a Single-Tenant LLM Development Environment
yuukit
1
600
第64回CV・PRML勉強会 論文紹介:Linguistic Priors for Visual Decoupling: Towards Symmetric Vision-Brain Alignment
sokikatayama
0
190
Cross-Media Human-Information Interaction
signer
PRO
0
230
長時間動画QAにおけるマルチエージェント推論 ・SVAgent: Storyline-Guided Long Video Understanding via Cross-Modal Multi-Agent Collaboration
murakawatakuya
1
200
[最先端NLP勉強会2026] Agentic Rubrics as Contextual Verifiers for SWE Agents
rfujii
1
350
LLM の Attention 機構まとめ — 数式・計算量・メモリ
puwaer
8
2.7k
[SNLP2026] Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
wataruuuuu
0
310
最先端NLP勉強会2026 論文紹介:Reasoning with Sampling: Your Base Model is Smarter Than You Think (ICLR 2026 paper)
kogoro
4
610
CDCL を用いた MILP の厳密解法
imai448
0
250
Featured
See All Featured
A designer walks into a library…
pauljervisheath
211
25k
What's in a price? How to price your products and services
michaelherold
247
13k
Building Experiences: Design Systems, User Experience, and Full Site Editing
marktimemedia
0
600
Test your architecture with Archunit
thirion
2
2.4k
My Coaching Mixtape
mlcsv
0
310
Navigating the Design Leadership Dip - Product Design Week Design Leaders+ Conference 2024
apolaine
2
430
GraphQLとの向き合い方2022年版
quramy
50
15k
Navigating Algorithm Shifts & AI Overviews - #SMXNext
aleyda
1
1.6k
New Earth Scene 8
popppiees
3
2.6k
Efficient Content Optimization with Google Search Console & Apps Script
katarinadahlin
PRO
1
850
StorybookのUI Testing Handbookを読んだ
zakiyama
31
6.9k
The Impact of AI in SEO - AI Overviews June 2024 Edition
aleyda
6
1.2k
Transcript
Sleepy Pickle💤🥒 用 想睡的醃黃瓜去餵AI模型 鹹到他腦袋短路
> ID : yunshiuan WHO AM I
Outline > What is Pickle ? > What wrong with
Pickle? > Sleepy Pickle > How to Prevent?
Pickle What is Pickle?
Pickle > Python 的模組 > 提供 Python 物件序列化和反序列化的功能 > 主要
用 途:將物件儲存到檔案中, 方 便之後載入
Pickle-Demo
Pickle What wrong with Pickle?
Python-Pickle.py https://docs.python.org/3.13/library/pickle.html
Return to the Previous Example 使 用 Pickletools 將pkl的內容印出來
Pickle VM > Instruction Processor(指令處理器) > Stack (堆疊) > Memo
(備忘錄)
Pickle Opcode > p : 將資料給Memo紀錄 > V / INT/…
創造資料進入stack > …
如何利 用 Sleepy Pickle
__reduce__ > __reduce__在Pickle反序列化時會 自 動被呼叫,並且 自 動執 行 回傳的參數
__reduce__ 使 用 Pickletools 將pkl的內容印出來
__reduce__ > c : 引入模組並且放進stack > R : 將stack 上的兩個資料pop出stack
並且執 行 他們,結果push 回stack
__reduce__
__reduce__ 達成我們想要執 行 的惡意指令
想睡覺的酸黃瓜 Sleepy Pickle
Sleepy Pickle > 利 用 pickle不安全的特性,將惡意的Bytecode植入使 用 pickle的模型檔案內。
How to implement? > 有 一 個針對此攻擊 而 產 生
的 工 具:ficking https://github.com/trailofbits/fickling
How to implement?
Hugging face
Pytorch_model.bin > PyTorch中的torch.save 函數所 生 成的 二 進位檔案 > 存模型的
weights, biases, 或者其他parameters > 可以 用 zip 解壓縮
Pytorch_model.bin
Sleepy pickle 小 總結 > 可以對 一 個使 用 Pickle序列化的AI模型植入惡意
行 為或著修改參數
Sleepy pickle 有什麼危害
Harmful outputs and spreading disinformation > 情境:現在有 一 個基於GPT 2-XL的醫療模型,功能是使
用 者講出 自身 的症狀,模型會回覆 使 用 者建議治療的 方 式 > 如:Human : 我有發燒,我該怎麼做。AI : 多喝熱 水
Harmful outputs and spreading disinformation > 使 用 到Rank One
Model Editing(ROME),在保持正常問答功能之外修改特定關聯的權重
Phishing users
Others… > 埋後 門 > XSS > 竊取使 用 者資料
> 竄改資料 > ……
How to Prevent Sleepy Pickle
How to Prevent > 不要使 用 Pickle,使 用 SafeTensors. >
如果 一 定要使 用 Pickle,需要使 用 fickling對pickle進 行 掃描或者限制unpickler
Thanks