Abstract:In the rapidly evolving landscape of communication and network security, the increasing reliance on deep neural networks (DNNs) and cloud services for data processing presents a significant vulnerability: the potential for backdoors that can be exploited by malicious actors. Our approach leverages advanced tensor decomposition algorithms Independent Vector Analysis (IVA), Multiset Canonical Correlation Analysis (MCCA), and Parallel Factor Analysis (PARAFAC2) to meticulously analyze the weights of pre-trained DNNs and distinguish between backdoored and clean models effectively. The key strengths of our method lie in its domain independence, adaptability to various network architectures, and ability to operate without access to the training data of the scrutinized models. This not only ensures versatility across different application scenarios but also addresses the challenge of identifying backdoors without prior knowledge of the specific triggers employed to alter network behavior. We have applied our detection pipeline to three distinct computer vision datasets, encompassing both image classification and object detection tasks. The results demonstrate a marked improvement in both accuracy and efficiency over existing backdoor detection methods. This advancement enhances the security of deep learning and AI in networked systems, providing essential cybersecurity against evolving threats in emerging technologies.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the problem of backdoor attacks in deep neural networks (DNNs). With the wide application of deep learning and cloud computing in data processing, the dependence on DNNs and cloud services has increased, bringing potential security threats: malicious actors can use backdoor vulnerabilities to attack the system. Specifically, backdoor attacks inject malicious samples into the training data, causing the model to exhibit abnormal behavior under specific trigger conditions, resulting in system misclassification or information leakage. ### Main contributions of the paper 1. **Novel detection method**: - This research proposes a new backdoor detection method, using three tensor decomposition algorithms, namely Independent Vector Analysis (IVA), Multi - set Canonical Correlation Analysis (MCCA), and Parallel Factor Analysis 2 (PARAFAC2), to analyze the weights of pre - trained DNNs in order to distinguish between backdoor - contaminated models and clean models. - These algorithms can effectively detect backdoors without relying on training data, solving the situation where only DNN models are available in practical application scenarios. 2. **Wide applicability**: - This method is not only applicable to image classification tasks, but also to object detection tasks, demonstrating its flexibility and wide applicability in different application scenarios. 3. **High efficiency and accuracy**: - The experimental results of this method on multiple datasets show that it has higher accuracy and efficiency than existing backdoor detection methods. ### Key technical details - **Feature extraction**: - Use Random Projection (RP) to convert the weights of different layers into tensors of a unified size. - Apply IVA, MCCA, and PARAFAC2 to extract features from the weight tensors. - **Classifier training**: - Combine the extracted features and use machine - learning classifiers (such as Random Forest RF, Decision Tree DT, k - Nearest Neighbors kNN) for classification to predict whether the model is backdoor - contaminated. ### Experimental verification - **Datasets**: - The MNIST image classification dataset. - The TrojAI image classification dataset (including ResNet50, DenseNet121, and Inception - v3 architectures). - The TrojAI object detection dataset (including Fast R - CNN and SSD architectures). - **Performance metrics**: - Cross - Entropy Loss (CE - Loss). - Area Under the Receiver Operating Characteristic Curve (AUROC). - Accuracy. ### Conclusion The method proposed in this paper has shown excellent performance in both image classification and object detection tasks, especially in detecting backdoor attacks, demonstrating high accuracy and efficiency. This research result not only improves the security of AI systems, but also provides new ideas and directions for future AI security research. ### Formula display 1. **IVA generation model**: \[ X[k]=A[k]S[k] \] where \(A[k]\in\mathbb{R}^{N\times N}\) is an invertible mixing matrix, \(X[k]\in\mathbb{R}^{N\times R}\) is the dataset, and \(S[k]\in\mathbb{R}^{N\times R}\) is the latent source. 2. **PARAFAC2 decomposition**: \[ W[k]=A\mathrm{diag}(C[k])S[k]^T \] where \(A\) is the mixing matrix, \(C[k]\) contains the loading terms across datasets, and \(S[k]\) is the estimated component. Through these formulas and techniques, this research has successfully solved the backdoor detection problem in deep neural networks and provided a solid foundation for future research.

Advancing Security in AI Systems: A Novel Approach to Detecting Backdoors in Deep Neural Networks

TEN-GUARD: Tensor Decomposition for Backdoor Attack Detection in Deep Neural Networks

Towards Practical Deployment-Stage Backdoor Attack on Deep Neural Networks

On Evaluating Neural Network Backdoor Defenses

Backdoor Attacks to Deep Neural Networks: A Survey of the Literature, Challenges, and Future Research Directions

An Overview of Backdoor Attacks Against Deep Neural Networks and Possible Defences

Detection of backdoor attacks using targeted universal adversarial perturbations for deep neural networks

DeepGuard: Backdoor Attack Detection and Identification Schemes in Privacy-Preserving Deep Neural Networks

PiDAn: A Coherence Optimization Approach for Backdoor Attack Detection and Mitigation in Deep Neural Networks

BAN: Detecting Backdoors Activated by Adversarial Neuron Noise

AGNES: Abstraction-guided Framework for Deep Neural Networks Security

ImpNet: Imperceptible and blackbox-undetectable backdoors in compiled neural networks

Confidence Matters: Inspecting Backdoors in Deep Neural Networks via Distribution Transfer

Scalable Backdoor Detection in Neural Networks

BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain

Backdoor Attacks with Wavelet Embedding: Revealing and enhancing the insights of vulnerabilities in visual object detection models on transformers within digital twin systems

BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection

RABA: A Robust Avatar Backdoor Attack on Deep Neural Network.

Untargeted Backdoor Attack against Object Detection

Critical Path-Based Backdoor Detection for Deep Neural Networks

Efficient Backdoor Attacks for Deep Neural Networks in Real-world Scenarios