What I did

• —Co-developed TRACE, a text-to-image steering framework triggering sparse-autoencoder steering only at high-risk denoising steps.
• —Cut attack success rate by nearly 50% on two test models at ~1% latency overhead, using only 48 MB of added weights.
• —Trained 50+ per-layer linear risk probes and selected the most predictable layer by AUPRC score and SAE reconstruction error (FVU).
• —Built the projection and mean-pooling, compressing activations 256× (512 KB → 2 KB), enabling a single-pass sweep across all layers.

Tools

PyTorch · scikit-learn · NumPy · LlamaGuard/XGuard · Hugging Face Diffusers