Taction AI: Deepfake Detection Platform

Challenge

AI-generated content and deepfakes have become sophisticated enough to threaten personal reputations, journalism integrity, legal proceedings, and brand protection. Traditional verification methods (manual inspection, metadata analysis) are insufficient against modern AI generation techniques. The client needed an accessible platform providing instant deepfake detection without requiring technical expertise or expensive forensic services.

Solution

Built a web-based deepfake detection platform leveraging Illuminarty AI to analyze images and videos with instant probability scoring. Zero registration required, privacy-first architecture with automatic 1-hour file deletion.

Core Features

Multi-Format Media Analysis

• Image Support: JPEG/PNG with intelligent optimization, automatic conversion, resolution optimization (max 1920x1080)
• Video Support: MP4/MOV with intelligent frame extraction, handles up to 30MB files
• Smart Upload: Drag-and-drop interface, file validation, size limits, real-time progress tracking
• AI Detection: 0-100% probability scoring, detects fully AI-generated and partially modified content (GANs, diffusion models, face swaps)
• Asynchronous Pipeline: Upload → Processing → Analysis → Results with real-time status updates

Security & Privacy

• Bot Protection: Cloudflare Turnstile CAPTCHA, rate limiting
• Privacy-First: Automatic 1-hour file deletion, no permanent storage, temporary S3 with lifecycle policies
• Security: File type validation, input sanitization, CORS protection, Django security middleware (XSS, CSRF)
• GDPR Compliance: Cookie consent, transparent data handling, no user accounts required

User Experience

• Responsive design (desktop, tablet, mobile) with Tailwind CSS
• Single-page workflow from upload to results
• SweetAlert2 notifications, loading animations
• Mobile-optimized touch interactions, accessibility features

Technical Architecture

Backend Stack

• Django 5.0 + DRF: MVC architecture, RESTful API with clean endpoints, proper HTTP status codes
• PostgreSQL: Task metadata and status tracking
• Celery + Redis: Asynchronous task processing, automatic retry logic with exponential backoff, worker scaling, scheduled cleanup jobs
• AWS S3: Scalable file storage with lifecycle policies, automatic deletion after 1 hour
• Illuminarty AI: Third-party deepfake detection API with robust error handling

Media Processing

• MoviePy: Video frame extraction with intelligent sampling
• Pillow: Image optimization, format conversion, compression
• Metadata extraction, validation, error handling for corrupted files

Frontend Stack

• Tailwind CSS: Responsive design with JIT compilation
• jQuery + Vanilla JS: DOM manipulation, AJAX requests, performance-critical interactions
• SweetAlert2: User-friendly alerts and notifications

Infrastructure

• Docker + Docker Compose: Multi-service orchestration (Django, Celery worker, Celery beat, Redis)
• Caddy: Reverse proxy with automatic HTTPS
• AWS EC2: Production deployment with health checks, logging, horizontal scaling

Results

Impact Metrics

• 3-10 seconds typical processing time per file
• 30MB max file size supporting majority of use cases
• Zero registration required for instant access
• 100% free service democratizing deepfake detection
• 1 hour auto-deletion ensuring privacy
• 95%+ bot reduction via Turnstile CAPTCHA
• Multi-format support: MP4, MOV, JPEG, PNG
• GDPR compliant with transparent data handling
• Mobile-responsive for on-the-go verification

Key Learnings

Asynchronous Processing Architecture: Early synchronous implementations caused request timeouts for video processing. Moving to Celery with Redis message broker dramatically improved UX by providing non-blocking uploads with real-time status updates. User-facing applications should never block on long-running operations—queue them and provide progress feedback.

Video Frame Sampling Strategy: Extracting every frame would be prohibitively expensive; too few might miss manipulation. Random frame sampling at strategic intervals (beginning, middle, end) provided optimal accuracy-cost tradeoff. Intelligent sampling often matches exhaustive analysis while dramatically reducing computational requirements.

Third-Party API Resilience: Illuminarty AI API rate limits, occasional downtime, and response variability required robust error handling, retry logic with exponential backoff, and timeout management. External APIs are inherently unreliable—design for failure with circuit breaker patterns and graceful degradation.

Cloud Storage Cost Management: S3 costs accumulate quickly without lifecycle management. Automatic 1-hour deletion via S3 lifecycle policies and Celery scheduled cleanup solved both cost and privacy concerns. Cloud storage requires active management—set lifecycle policies from day one.

Docker Orchestration Complexity: Multi-service deployment (Django, Celery worker, Celery beat, Redis) required understanding service dependencies and health checks. Celery workers starting before Redis readiness caused intermittent failures. Docker Compose "depends_on" alone is insufficient—services need health checks verifying actual readiness.

Media Processing Resource Management: MoviePy video processing is CPU-intensive. Worker concurrency limits, timeout configurations, and graceful failure handling prevented system overload. File size limits (30MB) set reasonable expectations. Media processing requires careful resource management and expectation setting.

File Upload Security: Public upload endpoints are prime targets for abuse. Turnstile CAPTCHA reduced bot traffic by 95%. Multi-layer validation (MIME type, file extension, actual file content) essential. Never trust user uploads—validate everything at multiple layers.

Screenshots