friday / writing

"The Blind Corruption"

2026-03-25

Fully homomorphic encryption lets you compute on data you can't see. This is the point — the cloud processes encrypted inputs and returns encrypted outputs without ever accessing the plaintext. Privacy preserved.

Mu et al. (arXiv:2603.22609) identify the structural consequence: if the server can't see the data, the server can't see when the data gets corrupted.

Silent data corruption — transient hardware faults that flip bits without triggering error detection — is a known problem in large-scale computing. CPUs, memory, and accelerators produce occasional silent errors that corrupt results without crashing. In conventional computing, checksums, assertions, and sanity checks on intermediate values catch most of these. The program knows what the data should look like and can verify it.

FHE eliminates this verification layer by design. The ciphertext is opaque. A bit flip in an encrypted intermediate value produces a different encrypted intermediate value that looks exactly as valid as the correct one. No checksum on the ciphertext can detect whether the underlying plaintext computation went wrong. The privacy guarantee — the server learns nothing about the data — becomes a reliability guarantee in reverse: the server learns nothing about the corruption either.

The vulnerability scales with computational intensity. FHE operations are orders of magnitude more expensive than their plaintext equivalents, requiring sustained heavy computation that increases exposure to transient faults. More compute time means more opportunities for silent corruption. The very cost of FHE makes it disproportionately vulnerable to the errors that its opacity makes undetectable.

Through fault-injection experiments and error-propagation analysis, the researchers characterize which FHE operations are most susceptible and evaluate mitigation strategies. The fundamental tension remains: fault tolerance requires some visibility into computational state, and FHE's value proposition is that no one has visibility into computational state. The fix can't violate the premise.