## TL;DR
AI-powered digital forensics detects manipulated media -- from deepfakes to doctored images. The forensic AI arms race pits generation against detection, with cryptographic provenance (C2PA) providing a proactive solution: cameras sign images at capture, creating verifiable chains of custody from sensor to screen.

## Core Explanation
Digital forensics AI: (1) Deepfake detection -- classify real vs. AI-generated. Methods: facial artifact detection (inconsistent eye reflections, asymmetric earrings, blending artifacts at face boundaries), physiological signal analysis (heart rate from facial color changes -- AI-generated faces lack natural pulse), frequency domain analysis (GANs leave characteristic fingerprints in the frequency spectrum); (2) Image tampering detection -- splice detection (identifying copy-paste boundaries via noise inconsistency), resampling detection (interpolation artifacts from scaling/rotation), and lighting inconsistency (shadow direction, color temperature mismatch); (3) Audio forensics -- ENF (Electrical Network Frequency) analysis -- power grid frequency fluctuations recorded in audio serve as timestamps. AI detects synthetic speech by analyzing unnatural prosody patterns; (4) Device identification -- camera sensor pattern noise (PRNU) uniquely identifies individual cameras.

## Detailed Analysis
C2PA (Coalition for Content Provenance and Authenticity): backed by Adobe, Microsoft, Google, Intel, Sony, BBC. Key innovation: cameras generate a cryptographic signature at capture (binding image hash + metadata + timestamp + device identity). Any subsequent edit adds a signed manifest recording what was changed. The final image carries the complete provenance chain. C2PA 2.1 (2025): supports selective disclosure (showing cropped image without revealing full original), and integration with social media platforms for automatic "content credentials" display. IEEE Signal Processing (2024) survey: no single forensic method is universal. Best practice: ensemble of complementary detectors (face analysis + frequency analysis + metadata verification). AI forensics admissibility: for evidence to be admissible in court, the forensic method must be scientifically validated (Daubert standard). Current AI forensic tools face challenges: (1) Black-box methods -- difficult to explain to juries; (2) Rapid evolution -- a detector trained today fails against tomorrow's generators.