Year: 2025
Type: article
Abstract: Neural network applications are becoming a critical component of Public Digital Service Platforms (PDSPs), where they support biometric verification, document classification, intent detection, fraud prevention, and personalized accessibility services.As governments increasingly integrate AI-driven modules into electronic identification, e-governance interfaces, and public-service automation, the security of these neural models becomes a foundational requirement for ensuring trust, reliability, and regulatory compliance.However, the growing trend toward client-side execution-where models operate on heterogeneous mobile devices, selfservice terminals, or partially offline environments-significantly expands the attack surface.Hardware variability, intermittent connectivity, and locally cached model artefacts enable a wide spectrum of adversarial behaviors, including the crafting of adversarial examples that mislead classifiers, targeted poisoning of federated or incremental training datasets, privacy leakage through membership inference or model inversion, and large-scale extraction of model decision boundaries using automated querying.This paper proposes a comprehensive defense-in-depth framework specifically designed for PDSP environments with untrusted or semi-trusted clients.The framework combines adversarially robust training, TRADES-style robustnessaccuracy balancing, certified randomized smoothing, poisoning-resilient aggregation № 12(53) 2025 120 mechanisms, differential privacy guarantees, and watermark-based extraction detection.Complementary runtime safeguards-such as anomaly monitoring, behavioral detectors, and adaptive policy control-ensure that security persists even under dynamic adversarial conditions.Experimental evaluation using an electronic identification (eID) workload that integrates facial verification and document OCR demonstrates that the proposed architecture substantially improves robustness, adversarial detection accuracy, and privacy preservation while maintaining usability and computational efficiency.Robust accuracy increased from 0.62 to 0.84, AUROC for adversarial detection reached 0.95, and privacy-preserving learning retained over 95% clean accuracy.The results confirm that multi-layered defenses form a practical and sustainable foundation for securing neural network applications in PDSP ecosystems.Future research will focus on causality-aware detection mechanisms, automated watermark auditing pipelines, and fairness-constrained adaptive policies to further enhance trust, transparency, and resilience in AI-enabled public digital services.
Source: Наука і техніка сьогодні
Institution: State University of Trade and Economics
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4416666554