What is the primary goal of a backdoor (trojan) attack, and how does it differ from an evasion attack?

• A.A backdoor attack poisons training data so the model misclassifies inputs containing a specific trigger at inference time; an evasion attack perturbs test inputs at inference time without modifying training data
• B.A backdoor attack always requires hardware access; an evasion attack is software-only; trojaned models underperform clean ones on standard benchmarks by several points, meaning leaderboard parity rules out an implanted backdoor
• C.A backdoor attack degrades all predictions equally; an evasion attack targets only specific classes; a backdoor trigger changes the model's file hash, meaning integrity monitoring on the checkpoint catches trojaned weights at rest
• D.A backdoor attack targets regression models; an evasion attack targets classifiers