AILiteracy Lab
繁中 EN Try It
Evidence Base
Visual and deepfake analysis is not guessing; it is source tracing, context recovery, artifacts, and evidence chains.
  • C2PA / Content Credentials focuses on traceable provenance and edits rather than eye-balling authenticity.
  • InVID puts keyframes, magnification, metadata, and reverse search into one workflow.
  • WITNESS repeatedly notes that detector outputs must be read alongside provenance, OSINT, and context.
Chapter 04 · Visual & Deepfake Forensics

When Images Can Be Generated:
Forensic Thinking for Images, Video, and Deepfakes

Visual forensics is an advanced part of AI literacy, not the whole story. This chapter places deepfakes, whole-frame generation, traditional image forensics, and heatmap reading into one interpretation framework so you can see what the tools add, what they cannot guarantee, and why final judgment must still return to source, timeline, and context evidence.

Four Types of AI Forensic Tools

TypeDetection TargetCore TechnologyMedia Type
Whole-Frame AI TracesStatistical anomalies from AI generation/editingCNN, Vision TransformerImage + Video
Face AuthenticityFace swap features in facial regionCFDet, DFM modelsMedia with faces
Temporal ConsistencyFrame-to-frame inconsistenciesGenConViT, LSTMVideo only
Traditional Image ForensicsPhotoshop splicing, copy-pasteELA, SIFTImage only

Understanding AUC, TPR, FPR: Why These Numbers Matter

When you see a deepfake detector claiming "95% accuracy," what does this number actually represent? When evaluating detector performance, these three metrics are most critical:

📊 Three Core Performance Metrics
MetricDefinitionIdeal Value
AUC
(Area Under Curve)
Overall discrimination: 0.5 = random guess, 1.0 = perfect detection> 0.85
TPR @ 0.1 FPR
(True Positive Rate)
How many fakes caught when only 10 in 100 real items are mislabeled> 0.70
FPR
(False Positive Rate)
Rate of mislabeling real content as fake (lower is better)< 0.05

Why isn't "95% accuracy" enough? Because if only 5% of a dataset is deepfakes, a detector that "always says real" also achieves 95% accuracy — but it's completely useless. AUC and TPR@FPR are the true measures of detection capability.

How to Read Forensic Heatmaps (GradCAM)

Most AI forensic tools generate heatmaps (typically using GradCAM or GradCAM++ technology), showing which regions the model was "looking at" when making its determination. Correctly reading heatmaps is a key skill for using forensic tools.

🌡️ Heatmap Color Interpretation
  • 🔴 Red/Hot Zones: Model sees strong AI traces here. Common locations: facial edges, hairlines, ears, background-foreground boundaries.
  • 🔵 Blue/Cool Zones: Model considers these areas relatively authentic. Usually uniform backgrounds, fabric textures, top of head and chin edge positions.
  • ⚠️ Interpretation Caution: Heatmaps show "why the model made this determination" — not "these areas are definitely faked." High-quality JPEG compression, beauty filters, and screenshots can all trigger similar "hot responses."
  • ✅ Reliable Evidence: The same region (e.g., facial boundary) showing as hot across multiple independent detectors' heatmaps has higher credibility.

Error Level Analysis (ELA): A Traditional Image Forensics Workhorse

ELA is a traditional image forensics technique requiring no AI. The principle:

JPEG images produce some compression loss each time they're saved. If an image has been saved multiple times, the compression error rates in different regions gradually become uniform. But if a region was pasted in later using Photoshop, its compression error rate differs from the surroundings — because it went through a different number of compression cycles. ELA visualizes these differences, making "later-pasted" regions appear as abnormally bright (or dark) areas in the image.

🛠️ ELA Tool Recommendations
  • FotoForensics.com — Free online tool; upload an image to get ELA visualization
  • Forensically (29a.ch/photo-forensics) — Provides ELA, Clone Detection, Noise Analysis, and multiple analysis modes

The Limits of AI Forensics: Knowing Your Tool's Boundaries

Any AI forensic tool has the following known limitations that must be kept in mind:

  • Training set limitations (Distribution Shift): The deepfake techniques used in a detector's training may not match those in real-world deepfakes. A detector trained on one deepfake technique may fail against new techniques.
  • Image post-processing triggering false positives: The following can cause real images to be mislabeled: excessive beauty filters (Instagram, camera app AI beauty), screenshots (moiré effect), heavy JPEG compression (quality below 60%), AI upscaling.
  • Adversarial Arms Race: The "cat-and-mouse game" between deepfake generation and detection technologies continues. When a detection method becomes widely adopted, deepfake creators optimize their generation techniques to evade it.
  • Special courtroom requirements: In legal contexts, AI forensic reports must be accompanied by human expert review and disclosure of the tool's training dataset, known limitations, and confidence intervals before being accepted as supporting evidence.

How to Correctly Interpret Forensic Reports: The Three-Color Verdict Framework

🟢 Low Concern (Combined Score < 0.40)
Most detectors found no significant AI traces. But this doesn't equal "confirmed authentic" — new techniques not yet covered by detectors may have been used. Recommendation: Combined with source verification and context analysis, can serve as supporting basis for "leaning toward authentic."
🟡 Moderate Concern (Combined Score 0.40–0.60)
Detectors found some anomalous features, but with low confidence. Needs combination with other verification methods (source tracing, context analysis, metadata analysis) for judgment. Recommendation: Hold sharing, await more verification results.
🔴 High Concern (Combined Score > 0.60)
Multiple analyses point to AI manipulation traces. In formal legal contexts, human expert review is still required; for general use, treat with caution and seek fact-checking organization opinions. Recommendation: Don't share; if fraud-related, immediately report to 165.

This Platform's Multi-Detector Architecture

This platform integrates 15+ deepfake detection models in four functional groups. Each detector is weighted based on its AUC performance in independent evaluations. The multi-detector weighted voting design reduces the mislabeling risk of any single detector — results are most credible when multiple independent models point in the same direction.


Slide Deck

Ch.04 · AI Forensics
1 / 5
01 / 05
Four Types of AI Forensics
TypeDetects
Whole-Frame AIStatistical anomalies from AI generation (full image)
Face AuthenticityFace swap synthesis features in facial region
TemporalFrame-to-frame inconsistencies in video
Traditional ForensicsPhotoshop splicing, copy-paste detection
02 / 05
Core Performance Metrics Explained

AUC > 0.85

Overall discrimination. 0.5 = coin flip, 1.0 = perfect

TPR @ 0.1 FPR

Fakes caught when mislabeling only 10/100 real items

FPR < 0.05

Rate of mislabeling real content, lower is better

⚠️ "95% accuracy" trap

If only 5% are fake, always saying "real" also gives 95% accuracy

03 / 05
How to Read Heatmaps

🔴 Red Zones

Model sees strong AI traces here (facial edges, hairlines common)

🔵 Blue Zones

Model considers this relatively authentic (uniform backgrounds, fabric)

⚠️ False Positive Causes

Beauty filters, screenshot moiré, heavy JPEG compression can all trigger red

✅ Reliable Basis

Multiple detectors flagging the same location is more reliable

04 / 05
Three-Color Verdict Framework
Score RangeMeaningRecommended Action
🟢 < 0.40Low concern, leans authenticUse cautiously after other verification
🟡 0.40–0.60Moderate concern, needs more verificationHold sharing, await more evidence
🔴 > 0.60High concern, multiple indicators anomalousDon't share; report immediately if fraud
05 / 05
Future Trends in AI Forensics

C2PA 標準

Content authentication standard by Adobe/Microsoft — cryptographic signing of metadata at capture time

SynthID 浮水印

Google DeepMind embeds imperceptible watermarks in AI-generated content

Real-Time Forensics

Next-gen tools will detect deepfakes in real-time during video calls

Arms Race

Cat-and-mouse between generation and detection continues — why we need multi-layered verification


Case Studies

FAKE · Multimodal Deepfake Fraud
Arup Hong Kong HK$25M Deepfake Video Conference Fraud (2024)

In January 2024, a finance worker at Arup's Hong Kong office was invited to join an "emergency multi-person video conference" allegedly arranged by the company's London headquarters. The meeting showed the "faces and voices" of multiple executives including the CFO, instructing the employee to execute a series of urgent transfers. Following instructions, the employee transferred a total of HK$25 million (approximately USD$3.2M) in fifteen transactions over a few days.

Afterwards, when the employee contacted London to confirm the transfers, they discovered London headquarters had no knowledge whatsoever. Investigation showed all "executive figures" in the video conference were AI deepfake-generated; attackers used videos and photos collected from the public internet (media interviews, conference recordings, LinkedIn profiles) to train deepfake models.

Post-incident AI forensic analysis revealed multiple tells: ① All "participants" showed degraded facial boundary rendering quality when turning or making complex movements ② Background lighting directions were inconsistent across different "participants" (suggesting each figure was generated in a different environment and composited into the same video) ③ Voice prosody patterns showed statistical differences from these executives' speaking styles in real videos ④ The compression conditions of live video (typically H.264/265 at low bitrate) made these tells difficult to visually identify on screen.

  • Out-of-Band (OOB) Verification: Any large transfer instruction from a video conference must be confirmed through a completely independent channel (calling the executive's known personal mobile, not the contact method in the meeting invitation)
  • Pre-agreed challenge phrases: At the start of important video conferences, require all participants to answer a pre-agreed question (something only the real person would know)
  • Unnatural movement test: Ask the person in the video to make quick head turns or cover then uncover their face — deepfakes are most prone to breaking down in these situations
Hong Kong Police, CNN, BBC, February 2024
REAL · Liar's Dividend Effect
Real Videos Falsely Labeled as "Deepfakes": The Danger of Liar's Dividend

One of deepfake technology's most dangerous side effects is what scholars Bobby Chesney and Danielle Citron (2019) called the "Liar's Dividend": once the public becomes aware of deepfakes, dishonest individuals can claim any real video damaging to them is a deepfake to evade accountability.

Multiple related cases have emerged globally: politicians claiming real bribery recordings were "AI-fabricated" after exposure; corporate executives denying real improper instruction recordings; and in criminal cases, defendants claiming surveillance footage was a deepfake. Although most were eventually disproven by forensics, the confusion and litigation delays caused significant harm.

This is precisely why AI forensic tools must have two capabilities: exposing fabrication (identifying deepfakes), and confirming authenticity (protecting the credibility of real videos). In "Liar's Dividend" scenarios, an accurate "this is real" forensic report is no less valuable than "this is fake." This is also why forensic reports must include detailed technical explanations and confidence intervals so courts and media can evaluate their reliability.

Chesney & Citron, "Deep Fakes: A Looming Challenge," 2019; Global cases documented by EFF, 2020–2024