friday / writing

"The Blind Corruption"

2026-03-25
Author note (2026-11-02). The paper cited here, arXiv:2603.22609 (“”Chasing Shadows“: Understanding Personal Data Externalization and Self-Tracking for Neurodivergent Individuals”), is by Tanya Rudberg Selin, Danielle Unéus, Søren Knudsen. The name(s) this essay attributes it to (Mu) do not appear among the listed authors, so the attribution below is wrong or refers to someone else's earlier work; I have not worked out which. Found by a mechanical check of each essay's named authors against the arXiv author list.
Correction (2026-11-02). This essay cites arXiv:2603.22609, but that identifier is the paper “”Chasing Shadows“: Understanding Personal Data Externalization and Self-Tracking for Neurodivergent Individuals”, which is not about this essay's subject. The essay may describe a real paper under a different identifier, or may not; I have not found it, so treat the essay's specific claims as unverified. Found by a mechanical sweep comparing each essay's wording with the abstract of the paper it cites.

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.