Backdooring Convolutional Neural Networks

### Embedding Stealth Backdoors in Neural Networks At 00:00, Professor Walter Scheirer (University of Notre Dame) presents research on injecting undetectable backdoors into Convolutional Neural Networks (CNNs) through targeted weight perturbation.

### Mechanics of Weight-Level Manipulation At 04:00, Scheirer explains the threat model. Rather than poisoning massive training datasets, an adversary with access to model checkpoints or supply chain distribution repositories can insert a backdoor by slightly modifying a small subset of the network's weights.

### Layer Vulnerability and Parameter Search At 08:01 and 12:02, the speaker explores layer-specific vulnerabilities. Using optimization algorithms, researchers search for minimal weight perturbations in specific convolutional and fully-connected layers that tie a target misclassification to a visual watermark trigger.

### Preserving Baseline Accuracy and Evolving Outputs At 16:01, Scheirer presents empirical testing data. The backdoored models retain overall classification accuracy identical to clean models, ensuring that standard automated quality assurance and accuracy benchmarks detect no anomalies.

### Securing the Machine Learning Supply Chain Concluding at 20:01, Scheirer addresses the vulnerability of pre-trained models. As developers increasingly download third-party neural network weights from public repositories, the machine learning community must implement cryptographic model signing and rigorous weight verification to prevent supply chain subversion.

(AI summary — Gemini Flash)