Added DPA_VMI_FGSM Transfer Attack

[SaxenaAmogh]

Summary

This PR adds an implementation of the DPA_VMI (Dual-Perspective Transformation & Variance-Tuned Momentum) transfer attack to the face recognition transfer attack framework.

Paper basis:

• "Enhancing the Transferability of Adversarial Attacks through Variance Tuning" (CVPR 2021), Wang et al.
• "Improving the Transferability of Adversarial Attacks on Face Recognition with Diverse Parameters Augmentation" (CVPR 2025)

Changes

• Implemented dpa_vmi() in core/transfer_attack_core.py
• Added TensorFlow implementation of Nesterov Accelerated Gradient lookahead and DPA hard-anchor projective perspective batch transformations
• Added VMI variance estimation across $N=8$ noisy neighbor states around candidate lookahead states (v = avg_neighbor_grad - cur_grad)
• Added Translation-Invariant (TI) depthwise 2D Gaussian kernel spatial smoothing (tf.nn.depthwise_conv2d)
• Adapted the attack objective from image classification to face verification by using the repository's existing embedding similarity loss (compute_embedding() + attack_loss())
• Added DPA_VMI to ALL_ATTACKS
• Added the corresponding output column ('DPA_VMI': 'dpa_vmi_path') 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 core optimization pipeline of VMI & DPA, including:
  • Nesterov lookahead momentum trajectory
  • DPA projective perspective transform batching
  • Neighborhood variance tuning estimation
  • Gaussian Conv2D spatial translation-invariance smoothing
  • Iterative adversarial update under $L_\infty$ constraint ($\epsilon = 0.062$)
• The primary algorithmic adaptation replaces 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
DPA_VMI	37.92	0.2101
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

DPA_VMI breakdown:

Goal	Breach Rate (%)	Impact Mean
Dodging	47.08	0.2652
Impersonation	28.75	0.1549