Chapter 35 — Further Reading

Synthetic media is a fast-moving, contested field, so anchor your reading on sources that last — peer-reviewed papers, the standards themselves, and tools you can actually run — and treat any single "detector" the way the chapter treats a single score. Start with one rigorous paper and the provenance standard, then keep the tools at hand for when a finding is challenged.

Foundations (🔬 / deeper)

• Dolhansky et al., "The DeepFake Detection Challenge (DFDC) Dataset" (2020). The canonical evidence for the generalization gap — top models near the low-80s on the public set, ~65% on unseen fakes. Read it to understand why a detector's headline accuracy is not its courtroom accuracy.
• Hany Farid, Photo Forensics (MIT Press). The rigorous foundation for media authentication — geometry, lighting, sensor and compression artifacts — by the co-creator of PhotoDNA. The deep version of Chapter 20's discipline.
• Rössler et al., "FaceForensics++" (ICCV 2019). The benchmark dataset and paper for face-manipulation detection — the reference point for how detectors are trained, tested, and why they so often fail to transfer to new generators.
• C2PA Technical Specification (c2pa.org). The provenance standard itself: claims, assertions, COSE/X.509 signing, and the JUMBF binding embedded in the JPEG APP11 / PNG caBX structures. When a Content Credential is contested, this is the source.
• NISTIR 8280, Face Recognition Vendor Test Part 3: Demographic Effects (2019). The primary evidence that model error rates vary across demographic groups by orders of magnitude — the bias half of the Daubert problem, in numbers you can cite.

Approachable explanations (everyone)

• Chesney & Citron, "Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security" (California Law Review, 2019). The article that named the liar's dividend; readable, and still the clearest treatment of synthetic media's corrosive effect on evidence and trust.
• Content Authenticity Initiative (contentauthenticity.org). The plain-language home of Content Credentials — what they assert, what they do not, and how to read provenance. Pairs directly with the c2patool you will actually run.
• WITNESS, "Prepare, Don't Panic" (witness.org). A practitioner-and-advocacy resource on synthetic media in real cases — grounded, non-hype, and strong on the human stakes the sixth theme keeps central.

In practice (💾 Recovery · 🔍 Examiner · 🛡️ IR · 📜 Legal)

• 🔍 c2patool (github.com/contentauth/c2patool) and Adobe Content Credentials Verify (contentcredentials.org/verify). Inspect and validate the C2PA manifest in a file — the provenance check the chapter's authentication battery depends on. Report present / valid / absent precisely; never let "absent" become "fake."
• 🔍 DeepFake-o-meter (UB Media Forensics Lab, Siwei Lyu) and the FaceForensics++ benchmarks. A research aggregator of detectors and the standard benchmark — use them to understand spread and the generalization gap, never as a single verdict.
• 🛡️ OpenAI Whisper, Tesseract OCR, and spaCy. The open enrichment stack behind the triage pipeline — ASR transcription/translation, OCR, and NER for entity and link analysis. Every output is a lead to verify against the source.
• 💾 ssdeep/TLSH and Python imagehash (aHash/dHash/pHash), with OpenCV. Fuzzy and perceptual hashing for near-duplicate collapse, known-image matching, and prioritizing which carved fragments to restore first (Chapter 7).
• 📜 SWGDE validation guidelines, the Sedona/EDRM materials on TAR (predictive coding), and the C2PA specification. The validation, eDiscovery-review, and provenance references that make AI-assisted work defensible — pair with FRE 901/702 in Appendix E.

Reference (this book)

• Appendix C — Tool Reference: c2patool, the OCR/ASR/NER enrichment stack, perceptual-hash tools, and the detector frameworks named here.
• Appendix B — Python Forensics Toolkit: the perceptual-hash (dHash) and triage helpers from this chapter, scriptable for batch work.
• Appendix E — Legal Frameworks Reference: Daubert/FRE 702, FRE 901 authentication, and the statutes referenced clinically here.
• Chapter 20 — Photo, Video, and Document Forensics, Chapter 21 — Timeline Analysis, and Chapter 25 / 27 / 28: the authentication, timeline, and legal/ethics threads this chapter weaves; revisit Case Study 1 (triage with discipline) and Case Study 2 (the number that evaporated).

Do, don't just read

• Verify a Content Credential with your own eyes. Take a Content-Credentialed sample image, run c2patool photo.jpg --detailed, and read the signer and assertions; then change one pixel and watch validation fail, and upload the file to a social platform to confirm the credential is stripped — proving why absence is silent.
• Build a two-step triage on a practice image. Hash-filter against the NSRL (Appendix J), then run perceptual-hash dedup with Python imagehash and confirm you can collapse a re-saved JPEG and its original into a single review item.
• Run a battery on a control. Score a known fake and a known-authentic clip with several methods; note the overlap and the spread, feel the generalization gap, and write the result as "consistent with synthesis / insufficient to determine," never as a single number.
• Pin a manifest. Draft a one-paragraph methods statement and a model-and-version manifest — tool, version, threshold, validation set, measured error rate — that would survive a Daubert challenge. Reproducibility is admissibility.

Next: Chapter 36 — The Forensic Toolkit: step back from technique to the instruments themselves — Autopsy and TSK, FTK and EnCase, Cellebrite, Volatility, Wireshark — how to choose, validate, and combine them, weighing the AI features now embedded in every suite against exactly the reliability standard this chapter set.