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.