eprint.iacr.orgpaperNormal-Jet Masking: A Retraining-Free Defense Against Cryptanalytic Extraction of Neural Networks
Normal-Jet Masking (NJM) is a post-training wrapper that defends piecewise-linear neural networks against cryptanalytic parameter extraction without retraining. It keeps private mask channels inactive inside a calibrated tangent-normal tube around protected data and mixes them into outputs outside it. On seven MLPs it cuts raw-logit evidence F1 by 53.8 to 93.1%, raises attacker query cost 1.4 to 9.3 times, and removes pure hard-label duals. CNN agreement stays at 0.9993 or higher.
TL;DR Proposes a retraining-free wrapper that masks critical-point evidence so attackers cannot cryptanalytically extract neural network parameters.
Also seen from another angle🧠 AI →
Technologies named here
- Normal-Jet Masking (NJM)
Techniques named here
- critical-point extraction
- Cryptanalytic extraction of neural networks
- cryptanalytic model extraction
- Normal-Jet Masking
Concepts named here
- Agreement
- Critical-point evidence
- hard-label dual
- Pure hard-label duals
- Query cost
- Tangent-normal tube
Related in AI Security
blog.talosintelligence.comtoolIntroducing CAIRN: Frontier tracking for AI-integrated malware (Cisco Talos)
github.comadvisoryFlowise MCP Security Bypass Enables RCE
github.comrepoYesWeHack Claude Kit
jfrog.comarticle