Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A61 in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

The Dataset Matters: Linking Image Memorability to Adversarial Robustness

Ehsan Mohammed^1,2, Willow Han^1,2 Elham Bagheri^1,2, Apurva Narayan¹, Yalda Mohsenzadeh^1,2; ¹Department of Computer Science, Western University, ON, Canada, ²Vector Institute for Artificial Intelligence, ON, Canada

Presenter: Ehsan Ur Rahman Mohammed

Adversarial robustness is usually studied as a property of the model or training procedure, but the dataset itself mayalsoshapevulnerability. Weinvestigatewhetherim- age memorability, a human-derived intrinsic property of images, predicts adversarial robustness. Using the over- lapping classes from MemCat and THINGS, we build a class-balanced dataset with memorability annotations and compare low- versus high-memorability images un- der matched training and test splits. We fine-tune ViT- B/16, EfficientNet-B0, and ResNet-50, then evaluate ro- bustness under FGSM-L2, PGD-L2, and CW-L2 attacks. Across architectures, high-memorability images are con- sistently more robust than low-memorability images, with thestrongestandmostfrequenteffectsappearingforViT and ResNet-50. Training on low-memorability images also increases overfitting, as captured by a composite overfitting index built from training dynamics. Finally, local geometry analyses show that high-memorability images occupy denser regions of the learned mani- fold, while low-memorability images lie in sparser, less- supported regions. Together, the results suggest that memorability is not just a cognitive property of images, but also a useful data-centric predictor of adversarial resilience.

Topic Area: Memory, Learning & Knowledge Structures

Extended Abstract: Full Text PDF