feat(attacks-specific): ✨ Update PGD attack logic with customizable parameters and edge-case fixes

Co-Authored-By: Lilith Autocommit <noreply@atlilith.com>

services/imajin-adversarial/service/src/attacks/pgd.py

@@ -26,6 +26,8 @@ def pgd_l_inf(
     steps: int,
     alpha: float | None = None,
     random_start: bool = True,
+    clamp_lo: float = 0.0,
+    clamp_hi: float = 1.0,
 ) -> torch.Tensor:
     """PGD attack in L-inf norm against a custom loss function.
 
@@ -34,23 +36,28 @@ def pgd_l_inf(
         x:            Clean input tensor, shape (1, C, H, W), values in [0, 1].
         loss_fn:      Callable (x_adv) → scalar tensor to MAXIMISE.
                       loss_fn must call model internally and return a scalar.
-        eps:          L-inf perturbation budget in [0, 1].
+        eps:          L-inf perturbation budget, in the same scale as x.
         steps:        Number of PGD iterations.
         alpha:        Per-step size (defaults to eps / 4).
         random_start: If True, start from a uniform random perturbation in [-eps, eps].
+        clamp_lo:     Lower bound for valid pixel values (default 0.0 for [0,1] inputs).
+        clamp_hi:     Upper bound for valid pixel values (default 1.0 for [0,1] inputs).
+                      Pass -1.0 / 1.0 for models expecting [-1, 1]-normalised input (e.g.
+                      ArcFace) — mismatching the scale clamps dark pixels to 0.0, causing
+                      visible 127-unit shifts.
 
     Returns:
-        x_adv:  Adversarial example, same shape as x, clamped to [0, 1].
+        x_adv:  Adversarial example, same shape as x, clamped to [clamp_lo, clamp_hi].
     """
     if alpha is None:
         alpha = eps / 4.0
 
-    model.train(False)  # inference mode (no .eval() to avoid hook false-positive)
+    model.train(False)
     x = x.detach()
 
     if random_start:
         x_adv = x + torch.empty_like(x).uniform_(-eps, eps)
-        x_adv = x_adv.clamp(0.0, 1.0).detach()
+        x_adv = x_adv.clamp(clamp_lo, clamp_hi).detach()
     else:
         x_adv = x.clone().detach()
 
@@ -63,7 +70,7 @@ def pgd_l_inf(
             x_adv = x_adv + alpha * grad.sign()
             # Project back into L-inf ball around x
             x_adv = torch.max(torch.min(x_adv, x + eps), x - eps)
-            x_adv = x_adv.clamp(0.0, 1.0)
+            x_adv = x_adv.clamp(clamp_lo, clamp_hi)
 
     return x_adv.detach()