Add ANDA transfer attack

[Pranjali-Nagpal]

Summary

This PR adds an implementation of the ANDA (Asymptotically Normal Distribution Learning) transfer attack (CVPR 2024) to the face recognition transfer attack framework.

Paper basis: "Strong Transferable Adversarial Attacks via Ensembled Asymptotically Normal Distribution Learning" (CVPR 2024), Fang et al."

Changes

• Implemented anda_attack() in core/transfer_attack_core.py
• Added TensorFlow implementation of the ANDA running statistics (ANDAStatistics)
• Added TensorFlow equivalent of the affine translation ensemble used by ANDA
• Adapted the attack objective from image classification to face verification by using the repository's existing embedding similarity loss (compute_embedding() + attack_loss())
• Added ANDA to ALL_ATTACKS
• Added the corresponding output column to ATTACK_COLS
• Connected the attack through run_attack()

Notes

• Only core/transfer_attack_core.py has been modified, following the Assignment 4 submission guidelines.
• The implementation preserves the original ANDA optimization pipeline, including:
  • translation ensemble
  • running gradient statistics
  • Gaussian distribution estimation
  • iterative adversarial update
• The only intended algorithmic adaptation is replacing the original classification loss with the repository's embedding similarity objective to support both impersonation and dodging attacks.

Evaluation Results (docs/subset_input_pairs.csv):

Attack	Breach Rate (%)	Impact Mean
ANDA	36.67	0.1992
SI_NI_FGSM	29.17	0.1755
MI_FGSM	26.67	0.1643
MI_ADMIX_DI_TI	24.17	0.1534
TI_FGSM	20.42	0.1313
PGD	16.67	0.1002

ANDA breakdown:

Goal	Breach Rate (%)	Impact Mean
Dodging	45.42	0.2543
Impersonation	27.92	0.1441