Real-Time Deepfake Audio/Video Detector · Hacking Project

Detect AI-generated or manipulated audio and video in real time and return an explainable authenticity score for calls, conferences, and identity verification.

**The problem**
Generative AI now produces convincing synthetic voice and video, fueling CEO-fraud/vishing, KYC and biometric-onboarding bypass, disinformation, and impersonation attacks that defeat human perception and legacy liveness checks. Financial and identity-verification systems are being actively targeted with real-time voice cloning and face-swap streams. Organizations lack a real-time detector that can flag synthetic media in calls, video conferences, and onboarding flows before fraud completes.

**What you'll build**
Build a real-time deepfake detection system that analyzes audio and video streams for generative artifacts — spectral and prosodic inconsistencies in voice, and blink/lip-sync/frequency/temporal artifacts and face-warping in video — using CNN/transformer classifiers and frequency-domain analysis. Train on public deepfake corpora (FaceForensics++, DFDC, ASVspoof), add liveness and physiological signals (rPPG), and expose a low-latency API/SDK that returns an authenticity score and explainability heatmap. Continuously update against new generative models to counter the adversarial arms race.

**Objective** — Detect AI-generated or manipulated audio and video in real time and return an explainable authenticity score for calls, conferences, and identity verification.

**Expected outcome** — A real-time detector (API + demo UI) that scores live/uploaded audio and video for synthetic manipulation, with benchmark accuracy metrics and artifact-visualization heatmaps.

**Difficulty** — Expert — deep computer-vision and audio-DSP expertise, real-time low-latency inference, and constant model-drift/adversarial pressure make robust detection genuinely research-hard.

**Primary category** — AI/ML Security; Secondary = Digital Forensics (media forensics), Fraud Detection, Identity Security

**Security domains** — Primary = AI Security; Secondary = Security Research, Digital Forensics, Fraud/Identity Security, Threat Intelligence

**Technical focus** — deepfake detection, audio spoofing/anti-spoofing, face-swap/manipulation detection, spectral analysis, rPPG liveness, CNN/vision-transformer classification, real-time inference, media forensics, adversarial robustness

**Core technologies** — Languages: Python, C++ (latency-critical paths); Frameworks/Libraries: PyTorch/TensorFlow, OpenCV, librosa, dlib/MediaPipe, torchaudio; Security Tools: media-forensics toolkits, liveness SDKs; Cloud/Infra: GPU inference (CUDA/TensorRT), Docker, WebRTC pipeline, edge/streaming; AI/Data: FaceForensics++, DFDC, Celeb-DF, ASVspoof, WaveFake datasets; Databases: object storage for media, PostgreSQL for scores/metadata

**Where it's valued** — Banking/fintech KYC and fraud, video-conferencing platforms, media/news integrity, government/elections, social media trust & safety, telecom anti-vishing

**Careers it builds toward** — AI Security Researcher, Computer Vision/ML Engineer, Media Forensics Analyst, Fraud Detection Engineer, Trust & Safety Engineer

**Framework alignment**
• MITRE ATLAS: AML.T0048 External Harms (impersonation/fraud via generative media), adversarial-example evasion of the detector (AML.T0043 Craft Adversarial Data, AML.T0015 Evade ML Model)
• NIST CSF 2.0: Detect (detection of malicious/synthetic content), Identify (fraud risk to identity verification), Protect (identity assurance controls)
• OWASP: N/A to Web Top 10; relevant to LLM Top 10 only where synthetic media feeds multimodal model abuse
• CIS Controls v8: Control 6 Access Control Management (identity assurance), Control 14 Security Awareness (impersonation defense) — otherwise limited direct mapping
• Standards note: aligns with emerging content-provenance standards (C2PA) as a complementary control

**Innovation** — Research-Level Project / Startup Potential — real-time, generalizable deepfake detection remains an open research problem with clear commercial demand in fraud/KYC, giving genuine startup and patent angles. Scores: Innovation 9/10, Startup 9/10, Enterprise-Readiness 6/10, Patent = Possible.