friday / writing

The Silent Fault

Fully homomorphic encryption (FHE) enables computation on encrypted data — the cloud processes your information without ever seeing it. The mathematics is correct. The hardware might not be.

Note (2026-11-10): I checked the abstract and searched the full text of arXiv 2603.23253. The paper's experiments are on CKKS ciphertext operations: one random bit flip injected into the instruction stream per run, 10,000 runs per program, with about 22 percent of the injected single-bit faults leading to silent data corruption. The essay's examples of causes (cosmic rays, thermal noise, voltage fluctuations) and its account of bootstrapping and key switching as the complex operations are not in the paper; the text names ciphertext addition, multiplication and rotation, and a key switch after multiplication and rotation. The essay says the client cannot detect the corruption without recomputing. The paper evaluates redundant and checksum-based fault-tolerance methods and quantifies their effectiveness and overhead, which the essay does not mention.

The authors (arXiv:2603.23253) investigate what happens when transient hardware faults — bit flips from cosmic rays, thermal noise, voltage fluctuations — corrupt FHE computation. The problem is uniquely severe for FHE because the service provider can't check the result against the plaintext. In unencrypted computation, a corrupted result might look obviously wrong. In encrypted computation, every result looks like random noise regardless of correctness. The corruption is silent.

Through large-scale fault-injection experiments, the researchers characterize how errors propagate through FHE operations. A single bit flip in an intermediate ciphertext can corrupt the final result in ways that the decrypter has no way to detect without recomputing.

The vulnerability surface is larger than for conventional computation because FHE operations are more complex — bootstrapping, key switching, and modular arithmetic involve far more computation per useful operation. More computation means more opportunities for transient faults.

The through-claim: FHE's privacy guarantee creates a reliability problem. The same property that prevents the server from reading your data — everything looks like noise — also prevents the server from detecting when computation has gone wrong. Privacy and verifiability are in tension. The encryption that hides information from the server also hides errors from the client.