Cyber Risks · Trail of Bits Blog

PCC: Bold step forward, not without flaws

Summary: Apple announced Private Cloud Compute (PCC), a design that moves compute-heavy ML workloads like large language models off battery-powered devices and into Apple's cloud, where data must be decrypted for processing because homomorphic encryption does not currently scale. Trail of Bits assesses the conventional security design as excellent, but notes that PCC's guarantees depend on whether Apple uses it for training or inference, that Apple owns the attestation keys, and that Apple's custom GPUs remain affected by the LeftoverLocals vulnerability while Nvidia's hardware is not.

Why it matters: The data is decrypted on Apple's server, so the privacy claim rests on an enclave-based design rather than end-to-end encryption - and whether PCC is ever used for training, which Apple has not clarified, determines whether user data is retained in model weights. An attacker who could exploit both the load balancer and a side channel would need nation-state-level resources, which sets the practical bar for who can reach that data.

Read the original at Trail of Bits Blog ↗

Topics
AI Security
Related