Chapter 26 — Further Reading

Grouped by the book's three citation tiers (see _style-bible.md §7). Tier 1 = verified canonical sources we stand behind. Tier 2 = real ideas/literatures attributed honestly without a pinned-down exact citation. Tier 3 = illustrative/constructed material used for teaching. Annotations say what each is good for and, where relevant, its limits.

Tier 1 — Verified canonical

  • National Research Council (National Academy of Sciences), Strengthening Forensic Science in the United States: A Path Forward (2009). The field's reckoning and the validity yardstick this chapter applies to image forensics. Read it for the principle that a method's courtroom weight should rest on demonstrated reliability — directly relevant to why documented photogrammetry fares differently from facial comparison or standalone ELA.

  • President's Council of Advisors on Science and Technology (PCAST), Forensic Science in Criminal Courts: Ensuring Scientific Validity of Feature-Comparison Methods (2016). Sharpens the validity question into foundational validity and the demand for measured error rates. Use it to understand why facial comparison and deepfake detection — methods without well-characterized error rates on realistic inputs — cannot yet support courtroom-conclusion-level testimony, even as they retain investigative value.

  • Federal Rules of Evidence, Rule 702, and Daubert v. Merrell Dow Pharmaceuticals (1993); Kumho Tire Co. v. Carmichael (1999). The admissibility gate (Chapter 5). Especially relevant to novel techniques — automated face recognition, machine-learning enhancement, deepfake detection — where the testability, error rate, and acceptance factors are the crux of any challenge.

  • The public record of the 2013 Boston Marathon bombing investigation (Case Study 26.1). Valuable for seeing large-scale video evidence used well (synchronizing many independent cameras, reading behavior across frames) and the documented harm of crowdsourced facial misidentification (the wrongful online naming of innocent people). A single event that teaches both the power and the peril of image-based identification.

  • The public record of the 2006 Reuters / Adnan Hajj photo-manipulation controversy (Case Study 26.2). A clean, widely-documented example of image authentication working: cloning artifacts (impossible repetition) in a manipulated news photograph, caught by content-consistency reasoning, leading the agency to withdraw the images. Good for teaching that "it's a photograph" never meant "authentic," long before deepfakes.

  • The Innocence Project (innocenceproject.org), case and policy record. Background for the wider validity/wrongful-conviction argument, and relevant to the documented risks of mistaken identification — including from imagery — and of overstated forensic testimony presented as certainty.

Tier 2 — Attributed, specifics unverified

  • The forensic-photogrammetry literature on reverse-projection height estimation. A real body of work establishes reverse projection as the most defensible method for estimating stature from surveillance imagery and documents its error sources (stand-point placement, posture, footwear, camera changes, image quality). We attribute the existence and consensus of this literature and its standard cautions without citing a specific paper; any applied estimate should rest on a sound reconstruction with a stated uncertainty range.

  • The image- and video-authentication / multimedia-forensics literature. A substantial research field covers compression and JPEG forensics, error level analysis and its well-documented limitations, photo-response non-uniformity (sensor "fingerprint") source-device identification, and physics-based consistency checks (lighting, shadows, reflections, perspective). We attribute these techniques and, in particular, the limits of ELA (its confounding by edges/text/texture and its failure on re-saved or non-JPEG images) as established knowledge in the field, without pinning a single citation.

  • Research on the limits and biases of automated facial recognition. A recognized literature documents that face-recognition accuracy varies with image quality and, notably, across demographic groups, and that such systems are appropriately used as lead generators requiring corroboration, not as proof. We attribute the existence and direction of these findings without a specific citation; the bias dimension connects to Chapter 31.

  • The deepfake-detection and synthetic-media research literature. A fast-moving field documents generation methods (face-swap, full synthesis, voice cloning, lip-sync alteration), detection approaches (artifact cues such as blink/lighting/blending inconsistencies; machine-learning classifiers), and the core difficulties (rapid generalization gaps, degradation under compression and re-sharing). We attribute the field's state and its central caution — that detection is real but unsettled and not yet validated to a courtroom-conclusion standard — without overstating any specific method's reliability.

  • Content-provenance standards (e.g., the C2PA "content credentials" effort) and authentication-at-capture. A real, active standardization effort defines a tamper-evident way to attach a signed record of a file's origin and edit history. We attribute the existence and the design goals of such standards, and the strategy of proving genuine media's lineage rather than detecting fakes after the fact, without representing them as yet universal or settled in court.

  • The EXIF metadata standard and the fragility of image/video metadata. EXIF and analogous container/ stream metadata can record device, settings, time, and sometimes GPS; the chapter's caution that such metadata is easily stripped (e.g., by social-media re-compression) and editable is established and attributed here in general terms.

Tier 3 — Illustrative / constructed

  • The Mill Creek cold case (the Case File, and the gas-station CCTV, doorbell camera, and alibi-video facts). All cold-case persons, footage, and the "internally inconsistent metadata" of the alibi video are constructed teaching material, used to practice stating identity evidence as "consistent with" (not "identified") and to model the metadata asymmetry (inconsistency undercuts; consistency only fails to exclude). Clearly fictional; the persons of interest are invented.

  • Figure 26.1 ("Two enhancements of the same blur") and the worked plate/character details. A constructed teaching example illustrating the line between admissible enhancement (revealing partial, real information) and fabrication (a machine-generated crisp plate). The specific "six-pixel" figure and the legible characters are illustrative, chosen to make the rule transparent — not reference values.

  • The reverse-projection and ELA ASCII diagrams (§26.2, §26.4). Constructed schematics, explicitly not to scale, intended to convey the geometry of reverse projection and the workflow of error level analysis. Real reconstructions and ELA maps carry exact figures and image data, not the simplified marks shown here.

  • The illustrative height ranges used throughout (e.g., "approximately 178–184 cm," the exclusion/overlap examples). Illustrative round numbers chosen to demonstrate the exclusion-vs-consistent-with distinction; real estimates derive from an actual reconstruction with its own uncertainty.

Where to go next in this book

  • For the digital-evidence foundations this chapter builds on — forensic imaging, hash values, write blockers, and the fragility of metadata and cell-site data — see Chapter 25.
  • For the admissibility gate that novel image techniques must pass (Daubert, FRE 702), see Chapter 5; for foundational validity and the error-rate demand, Chapter 6.
  • For the cognitive-bias dynamics behind crowdsourced and analyst misidentification (confirmation bias, context management, blind comparison), see Chapter 31.
  • For how an expert presents a hedged, "consistent-with" finding without overstating under cross-examination, see Chapter 30.
  • For the financial trail that follows this chapter in the cold case — motive, the insurance policies, and the renovation-book anomalies — see Chapter 27; and for the capstone assembly of every thread, including the imagery, Chapter 39.