What question did this study set out to answer?

To develop a framework that enhances ANN-to-SNN conversion for improved AI interpretability and defense mechanisms.

February 8, 20260 citationsOpen Access

Activation-Scaled ANN-to-SNN Conversion with SNN Guardrail: A Unified Framework for AI Interpretability, Hallucination Detection, Real-Time Adversarial Defense, and Neural Healing

Key Points

To develop a framework that enhances ANN-to-SNN conversion for improved AI interpretability and defense mechanisms.
Utilized Spiking Neural Networks for real-time analysis of traditional AI models.
Implemented neural healing techniques achieving a 22% success rate on TinyLlama.
Conducted experiments with Mistral-7B to discover model-size-dependent thresholds.
Developed a demo platform on HuggingFace Spaces for interactive exploration of the framework.
Achieved a 100% jailbreak detection rate across eight attack types.
Identified that larger models require lower detection thresholds for effective adversarial defense.
Established a universal threshold formula for detection based on activation levels.

Abstract

I present a unified framework that extends ANN-to-SNN conversion beyond efficiency optimization to enable novel AI interpretability analysis, real-time adversarial defense, and autonomous neural healing. My approach uses Spiking Neural Networks as "computational microscopes" to analyze black-box AI models. **v5 Updates:**- NEW: Neural Healing v4A - Multi-stage progressive healing achieving **22% success rate** on TinyLlama (1.1B)- NEW: Mistral-7B (7B) experiment - Model-size-dependent threshold discovery- NEW: HuggingFace Spaces v2.0 - Live 3-tab demo (Jailbreak/Healing/Hallucination)- NEW: Revised Scaling Law - Larger models need *lower* detection thresholds **v4 Results:**- SNN Guardrail: **100% jailbreak detection rate** (8/8 attack types)- Scaling Law Discovery: TTFS sensitivity increases with model size (GPT-2: +3.1, TinyLlama: +4.2) **Previous Results (v1-v3):**- Universal threshold formula: θ = 2.0 × max(activation)- 100% accuracy preservation with hippocampal hybrid architecture- GPT-2 attention TTFS analysis: +3.1 increase for meaningless inputs- Hallucination detection: AUC 0.75 with ensemble classifier- ViT-Base (86M params) validation with CIFAR-100 Key insight: "Don't just block the AI when it's about to fail -- help it recover." 🔗 Live Demo: https://huggingface.co/spaces/hafufu-stack/snn-guardrail Code: https://github.com/hafufu-stack/temporal-coding-simulation/tree/main/ann-to-snn-converter

KI fragen

Bookmark

View Full Paper