trojan attacks

**Trojan Attacks** on neural networks are **attacks that modify the model's weights or architecture to embed a hidden malicious behavior** — unlike data poisoning (which modifies training data), trojan attacks directly manipulate the model itself to insert a trigger-activated backdoor. **Trojan Attack Methods** - **TrojanNN**: Directly modify neuron weights to create a trojan trigger that activates a hidden behavior. - **Weight Perturbation**: Add small perturbations to model weights that are dormant on clean data but activate on trigger. - **Architecture Modification**: Insert small additional modules (hidden layers, neurons) that implement the trojan logic. - **Fine-Tuning Attack**: Fine-tune a pre-trained model on trojan data to embed the backdoor. **Why It Matters** - **Model Supply Chain**: Pre-trained models downloaded from public repositories could contain trojans. - **Harder to Detect**: Direct weight-level trojans may evade data-level detection methods. - **Verification**: Methods like MNTD (Meta Neural Trojan Detection) and Neural Cleanse detect trojan behavior. **Trojan Attacks** are **sabotaging the model directly** — manipulating weights or architecture to embed hidden malicious behaviors that activate on trigger inputs.