SMASH 2026 · Paper #13 · Result card
Privacy-Preserving Text Inference for Health-Misinformation Classification and Semantic Retrieval
All 1,080 encrypted predictions matched plaintext. Accuracy 0.944, macro-F1 0.896 on CoAID; SIMD packing made encrypted retrieval 26 times faster.
claims 6 of yoursfeatures hybrid (556-d)text title+contentSVD dims 300CKKS ring N 8,192seed 42
CoAID test (1,080 items)
0.944 / 0.896
accuracy / macro-F1, hybrid features (556-d); paper hybrid 0.971 / 0.950
Encrypted = plaintext
1,080 / 1,080
identical labels; max logit error 6.0e-06
Your claims flagged “fake”
3 / 6
scored without the server ever seeing the text
SIMD packing speed-up
26×
33.9 s → 1.28 s over 5,168 items (paper 42.6×)
From the abstract
We design a privacy-preserving text inference framework based on CKKS homomorphic encryption for health-misinformation classification and semantic retrieval. The hybrid classifier achieved an accuracy of 0.971 and a macro-F1 score of 0.950, with identical reported metrics under plaintext and ciphertext-based inference.
Evidence
Paper vs. your run
| Metric | Paper | Your run | |
|---|---|---|---|
| Acc / macro-F1: TF-IDF (paper) vs TF-IDF→SVD300 | 0.942 / 0.904 | 0.944 / 0.895 | same recipe |
| Acc / macro-F1: MPNet (paper) vs potion-base-8M (256-d) | 0.938 / 0.899 | 0.928 / 0.866 | no BioBERT block; MPNet replaced by a static embedder |
| Acc / macro-F1: Full hybrid (paper) vs TF-IDF→SVD + potion (556-d) | 0.971 / 0.95 | 0.944 / 0.896 | no BioBERT block; MPNet replaced by a static embedder |
| Test split | 1,081 samples (184 fake, 897 real) | 1,080 (183 fake) | deduplicated CoAID news, title+content |
| Plaintext vs ciphertext predictions | identical accuracy and F1 scores for every feature configuration | 1080/1080 labels identical | N = 8,192 |
| Per-query cost at N = 32,768 | approximately 122.4 ms and 2.133 MB | 149.2 ms, 2.24 MB | 556-d vs 1,324-d vector; different CPU |
| Retrieval naive → SIMD-packed | 221.049 s → 5.186 s (42.6×), 6.11 GB → 2.13 MB | 33.9 s → 1.28 s (26×), 0.68 GB → 0.50 MB | 5,168 items, 256-d, N = 8,192; naive extrapolated from 200 |
Try next
- Title-only text with false-cure claims — Switching to title-only text strips article body context and brings inputs closer to social-media posts. The model may rely heavily on fact-check phrasing patterns present only in the full content field, so this tests whether style cues or substance drive the fake signal.
{"text_field": "title", "features": "hybrid"} - Paper-exact ring size N=32768 — Running the full encrypted test split at N=32,768 matches the paper's CKKS parameters exactly and lets you measure whether the 149 ms latency holds over all 1080 items (currently capped at 400 for the bulk pass), giving a direct apples-to-apples cost comparison.
{"poly_modulus_degree": 32768} - TF-IDF only, no embedder — In this run TF-IDF alone (0.944/0.895) nearly matched the hybrid (0.944/0.896). Running features=tfidf in isolation confirms whether the embedder is genuinely redundant or whether its contribution is masked by the SVD dimensionality, and is the closest single-block match to the paper's TF-IDF ablation row.
{"features": "tfidf"}
Read this before sharing
- A small static embedder replaces BioBERT and MPNet.
- ISCC-SCT fingerprints and late-fusion retrieval are not reproduced.