Green Edge AI for Cyber-Physical Systems: Stream-Native Topological Isolation against Adversarial Attacks in Resource-Constrained IoT
Year: 2026
Type: book
Abstract: Legacy $O(N)$ and $O(\log N)$ machine learning architectures encounter severe thermodynamic and geometrical limitations within large-scale artificial intelligence of things (AIoT) deployments. The Event Horizon - Decoupled Quartic-Gated (EH-DQG) framework, driven by a Cholesky-Gated Momentum (CGM) engine, resolves the stability-plasticity dilemma in resource-constrained environments. The architecture isolates anomalies in a whitened topological phase space. Stream-native execution achieves a strict per-packet time and space complexity of $O(d^2)$. Empirical hardware profiling on proxy microcontrollers demonstrates a practical engineering trade-off. The deterministic engine utilizes a peak working memory footprint of 4,176 Bytes and maintains a core state dimension of 456 Bytes. Algorithmic execution yields an empirically verified 11.19-year hardware-bounded autonomy limit. Thermodynamic expenditure requires a marginal energy penalty compared to static legacy models. The framework mathematically bounds adversarial vulnerabilities despite the minor resource exchange. The restrictive hyper-ellipse reduces maximum achievable physical damage from 18.00 down to 0.90. High resilience against gradual poisoning and manifold-targeted evasion is consequently ensured. A density-driven phase shift mechanism guarantees instantaneous recovery from abrupt concept drift without succumbing to catastrophic deadlocks. The proposed architecture ultimately provides a mathematically bounded, scalable Green Edge AI foundation for mission-critical cyber-physical systems, including deep-space constellations and implantable biomedical devices.
Institutions: Walailak University, University of Indonesia
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204635540