Neuracrypt is not private¶
Authors: Nicholas Carlini; Sanjam Garg; Somesh Jha; Saeed Mahloujifar; Mohammad Mahmoody; Florian Tramer
Year: 2021
Venue: arXiv preprint
type: attack collection: sok-references status: auto
Links: arXiv:2108.07256
Summary¶
Auto-extracted summary (unreviewed)
The paper breaks NeuraCrypt, an instance-encoding scheme that encodes training images by passing patches through a random-weight neural network, adding positional encodings and permuting patches. A boosting attack that learns a contrastive patch similarity (about 80% accuracy), aggregates it into an image similarity (over 95%), solves bipartite matchings, recovers the per-image and global permutations, and finally extracts the secret transformation by gradient descent matches original to encoded images with 100% success on ImageNet and on the public NeuraCrypt Challenge (CheXpert) for every security parameter, in under 6 hours on a single machine.
Abstract¶
NeuraCrypt (Yara et al. arXiv 2021) is an algorithm that converts a sensitive dataset to an encoded dataset so that (1) it is still possible to train machine learning models on the encoded data, but (2) an adversary who has access only to the encoded dataset can not learn much about the original sensitive dataset. We break NeuraCrypt privacy claims, by perfectly solving the authors' public challenge, and by showing that NeuraCrypt does not satisfy the formal privacy definitions posed in the original paper. Our attack consists of a series of boosting steps that, coupled with various design flaws, turns a 1% attack advantage into a 100% complete break of the scheme.
Source: arxiv
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | complete break of NeuraCrypt's privacy claims with 100% image-to-encoding matching | llm | 0.85 |
| contribution | iterative boosting attack that turns a weak patch-similarity advantage into a full break | llm | 0.85 |
| contribution | recovery of per-image patch permutations and of the secret encoding transformation | llm | 0.85 |
| contribution | perfect solution of the public NeuraCrypt Challenge | llm | 0.85 |
| contribution | demonstration that NeuraCrypt fails the formal privacy definitions of the original paper | llm | 0.85 |
| schemes | NeuraCrypt instance encoding | llm | 0.85 |
| libraries | JAX | llm | 0.85 |
| models | NeuraCrypt random-weight encoder | llm | 0.85 |
| models | contrastive embedding networks (attack) | llm | 0.85 |
| datasets | ImageNet | llm | 0.85 |
| datasets | CheXpert (NeuraCrypt Challenge) | llm | 0.85 |
| tasks | privacy-preserving training on medical images | llm | 0.85 |
| tasks | attack on instance encoding | llm | 0.85 |
| security level | NeuraCrypt security parameter k = 2, 7, 15 (all broken) | llm | 0.85 |
| threat model | adversary with access to the encoded dataset (matching game between original and encoded images) | llm | 0.85 |
| interaction | other | llm | 0.85 |
| code url | https://github.com/yala/NeuraCrypt-Challenge (URL found in the full text) | regex | 0.50 |
| limitations | attack is dominated by a quadratic-time all-pairs comparison between original and encoded images | llm | 0.85 |
| key result | attack success (matching accuracy): 100% (ImageNet, 10,000 images, k = 2 / 7 / 15) | llm | 0.85 |
| key result | attack wall-clock time: under 6 hours (single machine, NeuraCrypt Challenge) | llm | 0.85 |
| key result | patch similarity accuracy: up to 80% (balanced matching/non-matching patch pairs) | llm | 0.85 |
| key result | image similarity accuracy: over 95% (before permutation recovery) | llm | 0.85 |
SoK reference [53]¶
N. Carlini, S. Garg, S. Jha, S. Mahloujifar, M. Mahmoody, and F. Tramer, “Neuracrypt is not private,” arXiv preprint arXiv:2108.07256, 2021.
Listed on the reference page of the site.
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- dataset: ImageNet 5 [p.2]
- technique: ReLU 7 [p.2]
- assumption: IND-CPA 2 [p.11]
- task: Medical / healthcare 7 [p.1]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/sok-references/053_neuracrypt-is-not-private.pdf |
sok-references | 17 | 216.7 KB | 64c241fe6954 |
Database key: neuracrypt-is-not-private · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set neuracrypt-is-not-private COLUMN VALUE