JPEG Steganalysis Detectors Scalable With Respect to Compression Quality

Presented at IS&T's international symposium on Electronic Imaging, Media Watermarking, Security, and Forensics track 2020, San Francisco, CA, January 26–30, 2020.

Transcript

1. JPEG Steganalysis Detectors Scalable With Respect to Compression Quality Yassine Yousfi, Jessica Fridrich 1 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
2. Introduction Steganography Covert communication when secret messages are hidden in ordinary looking cover objects Focus on image objects because they contain many indeterministic components: acquisition conditions, development, post-processing, editing, and even sharing Images can have multiple formats: JPEG, GIF, PNG, etc. Secret messages are embedded by introducing noise looking signal to the image Can be content adaptive, e.g. taking advantage of textured regions, edges, etc. 2 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
3. Example of steganography 3 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
4. Introduction Steganalysis Identify which objects are “stego” i.e. containing secret messages Usually built as binary detectors, trained in sand-boxed environments: a known steganographic scheme, known payload, and a known cover source 2 approaches: Feature based detectors End-to-end deep learning based detectors 4 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
5. Deep learning in steganalysis Detection error for stego algorithm HILL at 0.4 bits per pixel using different steganalysis detectors Q M F 2002 SPAM 2009 CDF 2010 SRM 2011 m axSRM 2014 m axSRM + RC 2016 YeNet 2017 SRNet 2018 10 20 30 40 50 Detection error Deep CNNs 5 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
6. Motivation Myths of JPEG steganalysis “When facing multiple quality factors, grouping multiple quality factors in one detector is a bad idea” because it (unnecessarily) increases cover source diversity “CNNs are more sensitive to cover source mismatch (CSM) than rich models” ALASKA challenge (Sep 2018 – March 2019) Winners [Yousfi et al. 2019] trained a detector for each quality factor [60–100 QFs] Cumbersome, extremely time and resource consuming 6 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
7. Rule of thumb for grouping JPEG quality factors Rule of thumb Let q(Q) be the standard JPEG quantization table indexed by quality factor Q. The segment [Qmin, Qmax] can be grouped in one detector without loss of performance if qkl(Qmin)/qkl(Qmax) 2, ∀ 0 ≤ k, l ≤ 7 7 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
8. Setup of experiments Datasets BOSSbase + BOWS2, 20,000 grayscale 256×256 TRN / VAL / TST = 10,000 BOWS2 + 4,000 / 1,000 / 5,000 BOSS ALASKA v1, 49,500 color images, 256×256, four stego algorithms, diverse processing TRN / VAL / TST = 42,500 / 3,500 / 3,500 Performance measures Minimum total detection error PE = 100 × minPFA 1 2 (PFA + PMD ) Missed detection PMD for false alarm 5% MD5 = 100 × PMD (PFA = 5%) Detectors DCTR + FLD ensemble [Holub 2014, Kodovsky 2011] SRNet (deep residual CNN) [Boroumand 2018] 8 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
9. Segments of JPEG quality factors Maximum loss in PE w.r.t. detectors dedicated to a single QF 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 .25 .47 .20 .89 .44 1.04 2.04 .33 1.09 .80 J-UNIWARD at 0.4 bpnzac, SRNet, BOSSbase + BOWS2 99–100 covered separately by Reverse JPEG Compatibility attack 9 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
10. BOSSbase + BOWS2 PE of multi-quality vs.dedicated detectors (DCTR + ensemble) 70 75 80 85 90 95 98 JPEG Quality factor 25 30 35 40 PE MultiQF Dedicated 70 75 80 85 90 95 98 JPEG Quality factor 20 25 35 40 PE J-UNIWARD 0.4 bpnzac UED 0.3 bpnzac 10 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
11. BOSSbase + BOWS2 PE of multi-quality vs. dedicated detectors (SRNet) 70 75 80 85 90 95 98 JPEG Quality factor 5 10 15 20 PE MultiQF Dedicated 70 75 80 85 90 95 98 JPEG Quality factor 1 3 5 7 9 11 PE J-UNIWARD 0.4 bpnzac UED 0.3 bpnzac 11 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
12. ALASKA PE and MD5 of multi-quality detectors vs. dedicated tile (256×256) detectors (SRNet) 70 75 80 85 90 95 98 JPEG Quality factor 5 10 15 20 PE MultiQF Dedicated 70 75 80 85 90 95 98 JPEG Quality factor 7 13 19 31 37 MD5 12 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
13. Quality factors 99 and 100 PE of multi-quality detectors vs. dedicated Reverse JPEG Compatibility detectors [Butora 2019] QF 99 100 J-UNIWARD, Payload 0.1 0.05 0.1 0.05 Dedicated 6.84 20.11 0.02 0.54 Trained on QF99–100 6.96 19.41 0.09 0.43 Reverse JPEG Compatibility attack is also scalable 13 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
14. Robustness w.r.t. mismatched custom quantization tables Questions we want to answer Which generalizes better, CNNs vs. rich models? What is better, multi-quality detector vs. detector trained on closest QF? 14 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
15. Measuring quantization table dissimilarity Definition Semi-metric for comparing quantization tables q and p: d2(q, p) = 7 k,l=0 1 (k + l)2 qkl − pkl qkl + pkl 2 15 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
16. Testing suitability of semi-metric 80 90 80 90 80 90 80 90 80 90 PE Quantization table dissimilarity Minimum dissimilarity coincides with best detector SRNet, 10 custom QTs, J-UNI 0.4 (top rows), UED 0.3 (bottom rows) 16 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
17. Robustness w.r.t. custom quantization tables Loss in PE (PDedicated E − PX E ), where X is the detector trained on the closest QF (solid) and trained on the QF range (dashed) 12 14 −20 −15 −5 0 Loss in PE 5 10 15 Quantization table dissimilarity (x100) 5 10 15 12 14 −20 −5 0 Loss in PE 5 10 15 Quantization table dissimilarity (x100) 5 10 15 J-UNIWARD 0.4 bpnzac UED 0.3 bpnzac SRNet and DCTR+FLD-ensemble 17 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality
18. Key take aways Experimental rule of thumb proposed for grouping JPEG quality factors Both CNNs and rich models can span a range of JPEG quality factors CNNs better generalize to custom quantization tables Use larger batch size when training CNN detectors on a range Very dissimilar non-standard quantization matrices still a challenge 18 / 18 JPEG Steganalysis Detectors Scalable With Respect to Compression Quality