Learning Weight Perturbations during Neural Network Training

A newly disclosed training method injects deliberate weight perturbations into neural networks via a specialized regularizer, securing the first layer so protections propagate across all subsequent layers, according to a technology listing that states the approach is patent pending at TRL 3 and available for licensing. The method is designed to defend against snooping, cryptanalytic, and other attacks while eliminating feature encryption to improve energy efficiency on battery-constrained mobile, edge, and IoT devices, and it also targets neural networks delivered through machine learning service APIs. The listing claims the security measures are embedded directly in training without significant performance loss, distinguishing it from external protection techniques.

This technology introduces a method for enhancing the security of neural networks by applying weight perturbations during training to prevent unauthorized access and attacks. Neural networks are increasingly deployed in mobile, edge, and IoT devices, as well as through machine learning service APIs, raising concerns about their vulnerability to various attacks such as snooping and cryptanalysis. Traditional protection methods often fail to secure all layers of the network comprehensively, leaving models at risk. This need for robust security measures to safeguard neural networks in diverse environments led to the development of a novel training approach that strengthens model integrity from the initial layer forward. This technology also helps eliminate feature encryption, thereby significantly improving energy-efficiency of battery-constrained mobile, edge, IoT devices. This technology employs a specialized regularizer during neural network training that injects deliberate weight perturbations, effectively disrupting the mathematical dependencies between layers. By targeting the first layer, this method propagates security protections across all subsequent layers, ensuring the entire model remains safeguarded against potential attacks. This approach differs from conventional techniques by embedding security features directly into the training process rather than relying on external protective measures. The core innovation lies in the way these controlled perturbations create unpredictable model behaviors that deter malicious access and manipulation while preserving the network's overall performance. It is particularly designed for environments with limited computational resources, such as mobile and edge devices, where conventional heavy security protocols may be impractical. Additionally, this method enhances the security of neural networks delivered as services via APIs, which are common targets for adversarial attacks. By integrating this approach during training, developers can create neural models that are inherently resistant to various threats, providing a valuable solution for securing sensitive machine learning applications in modern interconnected systems. Photo for reference only, not a depiction of the invention. • Comprehensive security: Protects all neural network layers by securing the first layer, ensuring holistic model protection. • Efficient integration: Embeds security measures directly within the training process without significant performance loss. • Resource-friendly: Suitable for deployment on mobile, edge, and IoT devices with limited computational capacity. • Enhanced protection for services: Secures neural networks accessed via APIs, mitigating risks associated with remote attacks. • Robust against diverse attack vectors: Effective against snooping, cryptanalytic, and other common neural network attacks. • Mobile and edge device neural network models requiring enhanced security. • Internet of Things IoT devices utilizing machine learning components vulnerable to external threats. • Machine learning services provided through APIs that require protection from adversarial attacks. • Any artificial intelligence systems where safeguarding neural network integrity is critical for functional reliability. Patent Pending TRL = 3 This technology is available for licensing.