AFIP's Visual Media Analysis division develops forensic methodology for detecting manipulated images, AI-generated photographs, and deepfake video. As generative models produce increasingly photorealistic output, detection requires analysis at the sub-pixel level.
Every generative model leaves a characteristic fingerprint — patterns in the frequency domain tracing back to the model's architecture. AFIP's GAN fingerprinting methodology identifies not only whether an image is synthetic, but which model family produced it.
This attribution capability has significant implications for provenance tracking and forensic investigation.
Deepfake video manipulates facial features while preserving the surrounding frame. Our facial consistency analysis examines boundary regions, temporal coherence across frames, and physiological plausibility of facial movements.
Our methodology detects inconsistencies in lighting direction, skin texture continuity, and micro-expression dynamics.
Digital media undergoes multiple compression cycles during distribution. Each step introduces artifacts that interact differently with synthetic and authentic content. AFIP's framework detects signatures of AI-generated content even after multiple re-encoding passes.