Visual Question Answering on TextVQA Adversarial (epsilon_a = 4/255)
17.2VQA AccuracyFARE
Evaluation results
| Method | Links | |
|---|---|---|
| FAREAdversarial Training Strength (epsilon_t)=4/255, ET3 Defense=true2026.03 | 17.2 | |
| CLIPET3 Defense=true2026.03 | 17.1 | |
| FAREAdversarial Training Strength (epsilon_t)=2/255, ET3 Defense=true2026.03 | 15.3 | |
| TeCoAAdversarial Training Strength (epsilon_t)=2/255, ET3 Defense=true2026.03 | 13.2 | |
| TeCoAAdversarial Training Strength (epsilon_t)=4/255, ET3 Defense=true2026.03 | 12.6 | |
| FAREAdversarial Training Strength (epsilon_t)=4/255, ET3 Defense=false2026.03 | 10.2 | |
| TeCoAAdversarial Training Strength (epsilon_t)=4/255, ET3 Defense=false2026.03 | 9.5 | |
| TeCoAAdversarial Training Strength (epsilon_t)=2/255, ET3 Defense=false2026.03 | 7.8 | |
| FAREAdversarial Training Strength (epsilon_t)=2/255, ET3 Defense=false2026.03 | 7.2 | |
| CLIPET3 Defense=false2026.03 | 0.2 |