Technical Evaluation of a Disruptive Approach in
summary
Hope4Sec OÜ’s HDA (homomorphic data analysis) uses a key‑dependent hash that preserves similarity so existing tools run unmodified; encrypted Fashion‑MNIST (delta=3 and 6) is fed to TensorFlow MLPs and scikit‑learn models with an overall validation ≈89% and a 0.9534 random forest accuracy that beats many benchmarks. Data size shrinks by about factor 3, training time falls ~18% and validation ~12% on a single‑board computer, and black‑box probing cannot map ciphertexts back to plaintexts, yielding preserved classification (round index ≈1).
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:07Homomorphic encryption
- 4:03Third-party AI risks
- 8:00Fully homomorphic encryption limitations
- 11:35Classical schemes drawbacks
- 14:28Ashborphic AI intro
- 17:18AI encryption requirements
- 20:14Ashborphic AI security
- 22:00Encrypted datasets
- 26:15Experimental setup
- 28:22Performance on cybersecurity data
- 32:20Fashion dataset with TensorFlow
- 35:35Results and comparison
- 37:04Security and efficiency
- 38:55Conclusion
- 39:20Q&A
related talks
45mSchemes, Privacy, and Security
37mOpen-source Machine Learning on Encrypted Data
33mHomomorphic Encryption in the Open Source World
44mIP Protection and Privacy in LLM: Leveraging Fully Homomorphic Encryption
40mTrust No One
30mData Driven Security
39mTrust Cards For AI
41mNon-traditional Encryption Problems And Solutions