AI Media Forensics — manual

AI Media Forensics Manual

Practical Techniques for Detecting Synthetic Media

Version
Last Update

📑 Table of Contents

1. Introduction

Artificial Intelligence has enabled the creation of hyper-realistic synthetic media — images, video, audio, and text that can convincingly mimic authentic content. While AI brings innovation in media production, it also introduces risks: misinformation campaigns, reputational attacks, political manipulation, and cyber-enabled fraud.

This manual is designed for journalists, investigators, analysts, and digital forensic professionals. It delivers:

  • A structured methodology for content verification.
  • A multi-phase workflow that scales from rapid screening to evidentiary forensics.
  • A toolkit of practical technologies aligned with OSINT and DFIR practices.
  • Best practices for documentation, reporting, and transparency.

2. Analysis Models

The detection process is divided into four escalating phases:

  1. Rapid Triage (Initial Screening) – Quick suspicion check.
  2. Preliminary Verification (Lightweight Checks) – OSINT-based fast validation.
  3. Structured Forensic Analysis (In-Depth Review) – Comprehensive forensic-grade methods.
  4. Peer Review & Validation (Cross-Check) – Independent replication to reduce bias.

3. Detection Domains

How to use this section: Each domain targets a distinct failure mode common to synthetic media. Treat domains as independent lines of evidence. A single red flag rarely proves anything; two or more from different domains justifies escalation.

3.1 Anatomy & Object Integrity

Objective: Detect biological or object construction errors introduced by generative AI.

Indicators:

  • Extra, missing, or fused fingers; malformed nails; symmetrical eyes without natural variation.
  • Teeth rendered as uniform blocks or inconsistent with gum lines.
  • Ears, earrings, or glasses distorted or asymmetrical.
  • Clothing, fabric, or accessories with warped stitching, inconsistent patterns, or impossible geometry.

Checks:

  • Zoom to 200–400% and scan hands, eyes, and teeth.
  • Look for repeated face patterns in group shots.
  • Compare mirrored body parts for natural asymmetry.

Tools: Forensically, magnifiers, reverse image search on cropped anomalies.

3.2 Geometry & Physics

Objective: Test whether light, perspective, and reflections obey physical laws.

Indicators:

  • Shadows inconsistent with light sources or each other.
  • Reflections missing in mirrors, water, or glass.
  • Vanishing points misaligned; horizon misplaced.
  • Object scale inconsistent with distance.

Checks:

  • Use SunCalc to validate shadow length vs. claimed time/place.
  • Draw vanishing lines to test perspective.
  • Inspect reflections for parity and content.

Tools: SunCalc, Google Earth/Street View, Forensically.

3.3 Metadata & Technical Fingerprints

Objective: Analyze embedded metadata and camera/device signatures.

Indicators:

  • Missing EXIF in photos that should contain it.
  • Impossible timestamps or GPS coordinates.
  • Software tags showing AI editors or generators.
  • Uniform synthetic noise lacking natural PRNU (Photo Response Non-Uniformity).

Checks:

  • Run ExifTool to review Make/Model, DateTimeOriginal, GPS fields.
  • Inspect compression signatures and quantization tables.
  • Apply Noiseprint for sensor fingerprinting.

Tools: ExifTool, FotoForensics, Noiseprint.

3.4 Voice & Audio

Objective: Identify synthetic patterns in speech or environmental sound.

Indicators:

  • Robotic cadence; unnatural prosody.
  • Missing breathing, mouth clicks, or ambient noise.
  • Spectrogram anomalies: clean high frequencies, banding.

Checks:

  • Inspect spectrograms for unnatural frequency bands.
  • Measure jitter/shimmer in Praat for vocal variation.
  • Compare lip sync to phonemes in video.

Tools: Audacity, Praat, Deepware Scanner, Intel FakeCatcher.

3.5 Contextual Consistency

Objective: Confirm claimed time, place, and environment.

Indicators:

  • Seasonal mismatch (snow vs. claimed summer).
  • Buildings or skylines inconsistent with stated location.
  • Weather contradicting meteorological records.

Checks:

  • Validate shadows and lighting with SunCalc.
  • Compare weather with Meteostat or OGIMET logs.
  • Cross-reference landmarks via Google Earth or Street View.

Tools: Meteostat, OGIMET, Google Earth.

3.6 Behavioral & Social Signals

Objective: Assess realism of group dynamics and human behavior.

Indicators:

  • Identical faces or clothing repeated in crowds.
  • People ignoring focal events (all gazes in wrong direction).
  • Uniform expressions or synchronized gestures.

Checks:

  • Run face clustering to detect duplicates.
  • Check gaze direction consistency.
  • Observe micro-expressions and natural motion.

Tools: InVID, Forensically.

3.7 Textual AI Fingerprints

Objective: Detect linguistic artifacts of AI-generated text.

Indicators:

  • Repetitive scaffolding or formulaic phrasing.
  • Fabricated citations or unverifiable facts.
  • Uniform sentence lengths and transitions.

Checks:

  • Run AI detectors on samples.
  • Perform stylometric comparison to known author texts.
  • Spot-check quotes and references.

Tools: GLTR, DetectGPT, HuggingFace Models, JStylo.

3.8 Provenance & Watermarking

Objective: Identify provenance credentials or embedded watermarks.

Indicators:

  • Valid C2PA signatures showing edit history.
  • Invisible watermarks indicating AI generation.

Checks:

  • Extract provenance JSON and verify signatures.
  • Run SynthID or watermark scanners where available.

Tools: C2PA, Adobe Content Credentials, Google SynthID.

3.9 AI-vs-AI Detection

Objective: Apply specialized AI detectors trained to spot generative content.

Indicators:

  • High detector confidence across multiple frames.
  • Consistent outputs from different models.

Checks:

  • Apply forensic CNNs (XceptionNet, FaceForensics++).
  • Compare results across multiple detectors.

Tools: FaceForensics++, DFDC models, XceptionNet-based classifiers.

3.10 Cross-Modal & Narrative Consistency

Objective: Ensure all media modalities align with the narrative.

Indicators:

  • Lip sync mismatch between audio and video.
  • Weather sounds inconsistent with visual conditions.
  • Narration contradicting imagery.

Checks:

  • Align timestamps across text, audio, and video.
  • Verify environment acoustics match visual context.
  • Map camera positions vs. scene constraints.

Tools: CrossCheck, SensityAI, Reality Defender.

4. Step-by-Step Procedures. Step-by-Step Procedures

This section provides an operational, reproducible workflow from first contact with a file/link to an evidence‑grade conclusion. It is organized into four phases. Each phase includes objectives, inputs, actions, tools, outputs, and escalation criteria.

4.0 Pre‑Flight: OPSEC & Chain of Custody (CoC)

Objective: Preserve evidentiary integrity and avoid contaminating artifacts.

Inputs: Source URL, file(s), claims (who/what/where/when), stakeholder urgency.

Actions:

  • Acquire original if possible (avoid platform‑compressed versions). Request raw files via secure channel.
  • Hash immediately:
    • Bash/macOS: shasum -a 256 <file>
    • PowerShell: Get-FileHash <file> -Algorithm SHA256
  • Snapshot context: copy URL, post ID, author handle, timestamps (include time zone), and a screenshot of the claim.
  • Workspace: operate on a copy; never re‑encode originals. Record tool names & versions.
  • Risk & scope: decide if this is routine verification or high‑stakes (elections, conflict, criminal case).

Output: Case record with IDs, hashes, source notes, and a plan for Phase 1.

4.1 Rapid Triage (Initial Screening)

Objective: Decide in seconds whether the material merits deeper checks.

Inputs: One image/video frame, short audio snippet, or text excerpt.

Actions (by media type):

  • Image/Video frame:
    • Anatomy & objects: hands, eyes, teeth, ears, accessories, signage, logos.
    • Physics: shadow direction/length, reflections, specular highlights; lighting continuity.
    • “Too perfect” test: cinematic composition, hyper‑clean surfaces, uniform faces.
  • Audio: listen for breath/pauses, monotone prosody, robotic shimmer at pitch changes.
  • Text: repetitive phrasing, encyclopedic tone, confident statements without sources.

Common red flags: extra/merged fingers; mismatched shadows; mirrored or unreadable micro‑text; cloned textures; lip‑sync oddities; identical smiles.

False positives: heavy denoise/HDR; professional retouching; platform recompression; staged marketing visuals.

Output: Triage code — Green (plausible), Amber (suspicious), Red (multiple anomalies). Amber/Red → Phase 2.

4.2 Preliminary Verification (Lightweight Checks)

Objective: Use fast OSINT & basic forensic tools to confirm or challenge authenticity.

Inputs: Original (preferred) or best‑quality copy; claimed time/place/context.

Tools (typical): Google/Bing/Yandex Images; InVID‑WeVerify; ExifTool; Forensically / FotoForensics; Noiseprint; SunCalc; Timeanddate/Meteostat/OGIMET; Google Earth/Street View.

Step‑by‑step:

  1. Reverse search (image/video):

    • If video, extract 4–12 keyframes (InVID → Keyframes or ffmpeg -i input.mp4 -vf fps=1 frames/f%04d.jpg).
    • Search the full image plus cropped regions (faces, signs, skyline). Try horizontal flip when relevant.
    • Compare hits: earlier appearances, different captions, stock/AI galleries.

  2. Metadata inspection (images/video/audio):

    • exiftool <file> → review Make/Model, Software, DateTimeOriginal, GPS*.
    • Red flags: missing EXIF in camera JPEGs, impossible timestamps, odd Software (generator), GPS contradicting claim.
    • Caveat: social sites often strip/alter EXIF.

  3. Basic pixel forensics (images):

    • ELA/Clone/Noise in Forensically/FotoForensics.
    • Red flags: isolated high ELA around inserted objects; tiled repeats; uniform noise where natural variation is expected.
    • Noiseprint/PRNU hint: lack of camera‑like noise structure can support suspicion.

  4. Context cross‑check (all media):

    • Place: landmark geometry in Google Earth/Street View; signage language & fonts.
    • Time/lighting: SunCalc — does shadow azimuth/elevation match claimed date/time/location?
    • Weather: compare precipitation/clouds/temperature with Timeanddate/Meteostat/OGIMET.

Evidence to capture: screenshots of reverse‑search results; EXIF dumps; ELA/Noise overlays; SunCalc and weather pages (PDFs or images).

Decision & escalation:

  • Converging authentic signals → document as provisionally authentic.
  • ≥2 independent inconsistencies → escalate to Phase 3.

4.3 Structured Forensic Analysis (In‑Depth Review)

Objective: Produce a defendable assessment using advanced methods across modalities.

Inputs: Highest‑quality media; claims; any prior investigative notes.

Modules & procedures:

A) Video Forensics

  • Frame extraction:
    • Constant rate: ffmpeg -i in.mp4 -vf fps=5 frames/f_%05d.jpg
    • Scene changes: ffmpeg -i in.mp4 -vf "select='gt(scene,0.5)'" -vsync vfr scenes/s_%05d.jpg
  • Temporal artifacts: look for warping/morphing around faces/hands; inconsistent motion blur; jitter on edges; rolling‑shutter realism during pans.
  • Optical flow/consistency: check for motion coherence of shadows/reflections across frames.

B) Audio Forensics

  • Spectrogram analysis (Audacity): View → Spectrogram; inspect harmonics, breath noise, plosives; spot copy‑paste bands.
  • Prosody/phonation (Praat): measure pitch (F0), jitter/shimmer; overly uniform patterns suggest synthesis.
  • Deepfake detectors: run Resemble Detect / Deepware; treat as supporting, not decisive.
  • Physiological cues: where applicable, evaluate biometric pulse cues (e.g., FakeCatcher‑style signals) with caution.

C) Text Stylometry

  • Establish a baseline from verified writings (if authorship is at issue).
  • Analyze with JStylo (function words, POS patterns, sentence length variance).
  • Cross‑check with GPTZero/DetectGPT/HuggingFace classifiers; corroborate with factual verification (quotes, sources, dates).

D) Contextual OSINT

  • Geolocation: skyline line‑drawing; terrain/river bends; sign typography; street furniture; license plates.
  • Chronology: construction timelines (bridges, towers), event schedules, transport GTFS feeds.
  • Remote sensing: Sentinel Hub/NASA Worldview for cloud cover, snow extent, wildfire smoke on claimed dates.

E) Provenance & Watermarking

  • C2PA/Content Credentials: inspect with compatible viewers; export the provenance JSON; verify signatures and edit history.
  • SynthID/Watermarks: where tooling is available, check invisible watermarks in images/audio/text; document limitations.

F) Model‑Specific Forensics (AI‑vs‑AI)

  • Apply forensic CNNs (e.g., XceptionNet/FaceForensics++/DFDC models) on images/frames; never as a sole indicator. Record model type, version, thresholds, and confusion risks.

G) PRNU / Camera Fingerprinting (expert option)

  • Extract sensor noise residuals; compare to a reference set of images from the purported device.
  • Caveats: recompression, denoise, and resizing degrade PRNU; treat as corroborative.

Outputs:

  • Annotated frames/spectrograms; tool outputs (versions, parameters); OSINT corroboration; a reasoned conclusion with probability language (see 4.5).

Escalation triggers: conflicting signals; high impact (elections, criminal proceedings); legal request for expert affidavit.

4.4 Peer Review & Validation (Cross‑Check)

Objective: Reduce bias and ensure reproducibility.

Process:

  • Prepare a neutral brief (facts, methods, outputs) avoiding leading language.
  • A second analyst replicates key steps (reverse search, EXIF, pixel/audio/text analysis, context checks) independently.
  • Compare findings; document agreements and discrepancies; if needed, seek a third expert or additional data (original file, higher resolution, longer cut).

Artifacts: replication log, checklist of reproduced results, change log of conclusions.

Outcome: consensus conclusion or documented divergence with rationale.

4.5 Decision & Reporting Framework

Probability bands (recommendation):

  • Very Low (≤20%) — unlikely AI‑generated.
  • Low (21–40%) — weak indicators; more data recommended.
  • Indeterminate (41–59%) — conflicting signals; seek originals or expert tests.
  • High (60–80%) — multiple independent indicators of AI/manipulation.
  • Very High (>80%) — strong, corroborated evidence across domains.

Language examples: “High likelihood of AI generation based on [A, B, C], with no contradicting evidence. Limitations: [X, Y].”

Minimum evidence for publication (suggested): ≥2 independent indicators from different domains or 1 strong forensic indicator + context contradiction.

4.6 Automation Recipes (Optional)

  • Batch EXIF export:
    • exiftool -csv -r -DateTimeOriginal -Make -Model -Software -GPS* <folder> > exif_report.csv
  • Batch keyframes:
    • ffmpeg -i in.mp4 -vf fps=1 out/frame_%05d.jpg
  • Scene change list:
    • ffmpeg -i in.mp4 -filter:v "select='gt(scene,0.4)',showinfo" -f null - 2> scenes.log

Tip: Log tool versions and parameters alongside outputs for reproducibility.

4.7 Case Log Template (suggested fields)

  • Case ID, Analyst, Date/Time (TZ), Source URL/ID, Acquisition method, File hashes (SHA‑256), Media type, Claimed context, Tools & versions, Steps performed, Findings per step, Indicators (pro/contra), Probability band, Peer reviewer, Final conclusion, Evidence archive location.

5. Best Practices

  • Two-signal principle: Never conclude based on one indicator.
  • Documentation: Maintain chain of custody (hashes, metadata, tool versions).
  • Probabilistic reporting: Use “high likelihood” instead of absolutes.
  • Continuous adaptation: Update methods every 6–12 months.
  • Automation: Integrate tools into scripted pipelines.
  • Crowdsourced verification: Collaborate with OSINT/fact-checking communities.

6. Analyst Toolkit (2025)

  • Images: Forensically, FotoForensics, ExifTool, Noiseprint.
  • Video: InVID, ffmpeg, FakeCatcher.
  • Audio: Praat, Audacity, Resemble Detect, Deepware Scanner.
  • Text: GPTZero, DetectGPT, JStylo, HuggingFace classifiers.
  • Provenance: C2PA tools, Adobe Content Credentials, SynthID.
  • Context: Google Earth, Sentinel Hub, Meteostat, NASA Worldview.
  • AI-forensics: XceptionNet, FaceForensics++, DFDC models.

7. Strategic Outlook

  • Evolving AI: Generation models are rapidly improving, masking older flaws.
  • Future of detection: Watermarking, provenance standards, and blockchain-based verification will be critical.
  • Present reality: Only a hybrid approach (intuition + OSINT + forensics + AI detectors + provenance tools) can sustain investigative integrity.

Appendix A: Domain → Tools Matrix

Detection Domain Techniques Tools (Open-Source / Free)
Visual Forensics Identify anomalies: hands, eyes, teeth, reflections, shadows Forensically, Deepware Scanner, GIMP
Metadata & File Integrity Extract & analyze EXIF, XMP, hashes, signatures ExifTool, Mat2, Hashdeep
Error Level & Compression Analysis ELA, JPEG ghost detection, noiseprint mismatch FotoForensics, Noiseprint
Reverse Image/Video Search Reverse search images/videos for provenance InVID-WeVerify, Yandex Images, TinEye
Audio Forensics Spectrograms, waveform anomalies, deepfake audio classifiers Sonic Visualiser, Praat, FakeCatcher (Intel)
Textual Stylometry Stylometry, linguistic patterns, AI-text probability detectors GLTR, DetectGPT, HuggingFace Transformers
Context & OSINT Cross-Verification Cross-check geography, time, weather, events OSINT Framework, Wayback Machine, Bellingcat Tools
Network & Source Traceability Trace network origins, domains, C2PA provenance WhoisXML API, Maltego CE, RiskIQ, C2PA
Cross-Modal Consistency Check if narrative matches across modalities CrossCheck, SensityAI, Reality Defender
Automation & Pipelines Automate via pipelines, ML models, SIEM/XDR integrations Apache Tika, HuggingFace pipelines, Python, MISP, Sigma rules

Appendix B: Automation Snippets & Field Kit

Image & Metadata

  • Extract metadata (all files in folder to CSV): 
    exiftool -csv -r folder/ > exif_report.csv
  • Strip metadata for sharing (privacy): 
    mat2 file.jpg

Video Analysis

  • Extract 1 frame per second: 
    ffmpeg -i video.mp4 -vf fps=1 frames/out_%04d.jpg
  • Extract scene changes: 
    ffmpeg -i video.mp4 -filter:v "select='gt(scene,0.4)',showinfo" -vsync vfr scenes/out_%04d.jpg
  • Get video codec/container info: 
    mediainfo video.mp4

Audio Analysis

  • Convert to WAV for spectrograms: 
    ffmpeg -i input.mp4 -vn -acodec pcm_s16le output.wav
  • Generate spectrogram (SoX): 
    sox output.wav -n spectrogram -o spectro.png

Text Analysis

  • Detect AI-like text probability (DetectGPT): 
    from detectgpt import DetectGPT
model = DetectGPT()
score = model.score_text("sample text")
print(score)
  • Check perplexity with GPT-2 LM (HuggingFace): 
    from transformers import GPT2LMHeadModel, GPT2TokenizerFast
import torch

model = GPT2LMHeadModel.from_pretrained("gpt2")
tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")

text = "sample text"
encodings = tokenizer(text, return_tensors="pt")
max_length = model.config.n_positions
stride = 512
nlls = []
for i in range(0, encodings.input_ids.size(1), stride):
    begin_loc = max(i + stride - max_length, 0)
    end_loc = i + stride
    trg_len = end_loc - i
    input_ids = encodings.input_ids[:, begin_loc:end_loc]
    target_ids = input_ids.clone()
    target_ids[:, :-trg_len] = -100

    with torch.no_grad():
        outputs = model(input_ids, labels=target_ids)
        nlls.append(outputs.loss * trg_len)

ppl = torch.exp(torch.stack(nlls).sum() / end_loc)
print(ppl.item())

Networking & Provenance

  • WHOIS lookup (Linux): 
    whois example.com
  • Get SSL/TLS certificate info: 
    echo | openssl s_client -connect example.com:443 -servername example.com 2>/dev/null | openssl x509 -noout -dates -issuer -subject

Workflow Helpers

  • Batch hash files in folder:

    sha256sum * > hashes.txt

  • Create case log template (Markdown):

    # Case Log
- Case ID:
- Analyst:
- Date/Time (TZ):
- Source URL/ID:
- File Hashes (SHA-256):
- Media Type:
- Claimed Context:
- Tools & Versions:
- Findings:
- Indicators:
- Probability Band:
- Peer Reviewer:
- Final Conclusion:

    🔖 Credits

Maintained by Oryon +OSINT360 GPT.
This document is part of the Cyber Intelligence Toolkit project.

Themes