Evaluating Robustness of Multimodal Models Against Adversarial Perturbations
Upload an image to generate the adversarial image and caption using the APGD/SAIF algorithm.
Upload Image
Drop Image Here
- or -
Click to Upload
Adversarial Attack Algorithm
Epsilon (max perturbation, 0-255 scale)
↺
1
255
Sparsity (L1 norm of the perturbation, for SAIF only)
↺
0
10000
Number of Iterations
↺
1
100
Generate Captions
Original Image
Generated Original Caption
Perturbation (10x magnified)
Adversarial Image
Generated Adversarial Caption