DeepShield AI Forensics
Current deepfake detectors rely heavily on spatial convolutions, making them highly vulnerable to simple post-processing attacks like JPEG compression, resizing, or minor noise additions. I needed a system that could detect fakes even after they'd been laundered through social media compressions.
The Solution
I built a dual-pipeline forensic engine. It first extracts Error Level Analysis (ELA) and high-frequency Discrete Cosine Transform (DCT) artifacts, isolating compression discrepancies. It then feeds these artifacts into a Vision Transformer (ViT-Base-16) to classify the image. I also integrated an FGSM adversarial noise injection pipeline to pre-emptively immunize photos against facial cloning.
- Combined custom digital forensics algorithms (ELA, DCT frequency) with a pre-trained Vision Transformer (ViT-Base-16) model
- Built custom Grad-CAM visualization hooks to output pixel heatmaps of flagged anomaly areas
- Designed a dual-pass verification system to verify fakes that easily bypass standard CNN model boundaries
- Implemented a prototype FGSM noise-injector script to immunize original photos before web upload
System Pipeline
Engineering Trade-offs
Why?I opted for a Vision Transformer over standard CNNs (like ResNet) because ViTs capture global context better, despite the tradeoff of higher inference latency and memory requirements. I decoupled the Grad-CAM generation to run asynchronously to keep the main classification thread fast.
Proven Impact
Achieved a 97.5% classification accuracy on unseen, highly compressed GAN holdout sets. Reduced false-positive rates on heavily artifacted social media images by 34% compared to standard spatial detectors.