WatchDAWG

The problem. Generative AI can now produce images realistic enough to fuel misinformation, impersonation, and fraud at scale. But almost every detector gets evaluated on clean, untouched images — and real fakes don't stay clean. By the time a synthetic image reaches a real platform, it's usually been compressed, resized, filtered, or cropped, whether by the person spreading it or just by the platform itself. A detector that looks great in a lab benchmark can fall apart the moment it meets a real re-uploaded image.

The insight. We proved this ourselves before fixing it. We trained a detector the "normal" way — clean images only — and it hit 96% accuracy on clean test data. Then we threw the same required real-world transformations at it: JPEG compression, blur, resizing, noise, color shifts, cropping. Under heavy blur or aggressive downscaling, accuracy collapsed to 63% — barely better than a coin flip. Robustness isn't something a detector picks up for free. It has to be trained in on purpose.

The solution. WatchDAWG pairs a frozen, pretrained CLIP image encoder (so we're not training a vision model from scratch on limited data) with a small trainable classifier head, deliberately trained on corrupted versions of every image — the same 6 transformations the challenge requires, applied randomly during training. The whole model is 151M parameters, comfortably under the 2B limit, and only ~132K of those parameters are ever actually trained.

The proof. Training on corrupted images didn't just help — it transformed the model's worst-case behavior:

Worst-case accuracy (heaviest downscaling): 63% → 83%
Average accuracy across all 14 required transform/strength combinations: 83% → 91%
Cost: clean-image accuracy dropped by barely 1 point

The potential. This is a small, deployable proof of concept for something platforms genuinely need: a detector that doesn't just perform in a controlled benchmark, but holds up against images that have already been through the real distribution pipeline — reposted, recompressed, resized a hundred times before anyone official ever looks at them. And because it's small and cheap to run, it's a realistic building block for actual moderation infrastructure, not just a research demo