Deepfake & Synthetic-Media Analyzer · Hacking Project

Detect and explain synthetic or manipulated media across image, audio, and video with authenticity scoring and provenance checks.

**The problem**
Generative AI now produces convincing synthetic images, audio, and video used for fraud, CEO-voice scams, non-consensual imagery, and disinformation, and humans cannot reliably distinguish them. Enterprises and platforms lack accessible forensic tooling to flag manipulated media before it causes financial or reputational harm. This threat is accelerating faster than detection capability and provenance standards.

**What you'll build**
Build a multi-modal detection tool that analyzes images/audio/video for synthesis artifacts using deep-learning classifiers (CNN/transformer detectors), frequency-domain and physiological cues (e.g., blink/heart-rate rPPG for video, spectral artifacts for audio), and provenance signals (C2PA/Content Credentials, metadata forensics). Output an explainable authenticity score with highlighted manipulated regions and a chain-of-custody report.

**Objective** — Detect and explain synthetic or manipulated media across image, audio, and video with authenticity scoring and provenance checks.

**Expected outcome** — A working analyzer (API + UI) that ingests media and returns an authenticity verdict with artifact heatmaps, provenance status, and a forensic report.

**Difficulty** — Advanced — multi-modal deep learning, signal/media forensics, adversarial robustness, and provenance standards.

**Primary category** — Digital Forensics / Media Forensics; Secondary = AI/ML Security, Fraud Detection

**Security domains** — Primary = Digital Forensics; Secondary = AI Security, Threat Intelligence, Security Research, Incident Response

**Technical focus** — Deepfake detection, GAN/diffusion-artifact analysis, frequency-domain forensics, audio spoof detection, video physiological analysis (rPPG/blink), metadata/EXIF forensics, C2PA provenance, explainable AI, adversarial robustness

**Core technologies** — Languages: Python; Frameworks/Libraries: PyTorch/TensorFlow, OpenCV, librosa (audio), face/landmark libs, Grad-CAM (explainability); Security Tools: ExifTool, C2PA/Content Credentials SDK, media-forensics toolkits; Cloud/Infra: GPU compute, Docker, cloud inference; AI/Data: pretrained deepfake detectors, spectrogram models, datasets (FaceForensics++, ASVspoof); Databases: PostgreSQL, object storage

**Where it's valued** — Finance/anti-fraud, media/journalism, platforms/trust & safety, government/defense, legal/e-discovery, insurance

**Careers it builds toward** — Digital Forensics Analyst, AI Security Researcher, Media Forensics Specialist, Fraud Investigator, Threat Researcher

**Framework alignment**
• MITRE ATLAS: primary framework — adversarial use of generative models; detection counters GenAI-enabled impersonation and fraud. MITRE ATT&CK: Impersonation (T1656) in fraud contexts.
• NIST CSF 2.0: Detect (manipulated-media detection), Identify (threat assessment), Respond (forensic reporting)
• OWASP: LLM/GenAI Top 10 — synthetic-media misuse (LLM09 misinformation, deepfake abuse); N/A for classic web categories
• CIS Controls v8: Largely N/A; Control 8 (audit/chain-of-custody) applies to forensic integrity

**Innovation** — Research-Level Project — multi-modal detection with provenance integration is at the R&D frontier with clear enterprise/anti-fraud demand and patentable techniques. Scores: Innovation 8/10, Startup 8/10, Enterprise-Readiness 6/10, Patent = Possible.