refactor(membership-inference): migrate to new attack architecture

[asim29]

Summary

Two files updated (lira.py, dp_sgd.py) with targeted bug fixes and cleanup to align both with the base classes landed in PR 2. Integration and unit tests added.

Stacked on: pr/4.1-distribution-inference

---

What changed and why

attacks/lira.py

Change	Detail
experiment_id → exp_id	Renames the constructor parameter and super().__init__() call to match the base class attribute name established in PR 2.
Remove __get_logits, __normalize_logits	Both were dead code. __get_logits was never called in attack(); __normalize_logits was replaced inline.
Simplify attack() logit collection	Pre-allocates all_logits as list[list[np.ndarray]] with a known length instead of growing it lazily. Targets collected with .numpy() directly (they were already on CPU after DataLoader).
Simplify in_out_labels construction	Replaced the per-shadow loop with a single np.array([...], dtype=bool) list comprehension.
Remove copy.deepcopy in __get_log_logits	The softmax is computed on a per-model slice (pred_logits[model_idx]), so modifying probs in-place never affects the original array.
Return type annotation	attack() now declares -> dict[str, np.ndarray] as required by the ABC.
Minor: lira_offline_preds, _	Second return value from __lira_offline was being silently overwriting true_labels; changed to _.

defenses/dp_sgd.py

Change	Detail
Unbound loss variable	loss.item() in the end-of-epoch print was referencing the loop variable after the loop might have zero iterations. Fixed by introducing last_loss: torch.Tensor | None = None updated each iteration, guarded by if last_loss is not None.
self.trainloader → self.train_loader	The Opacus make_private() return value was being assigned to a new attribute self.trainloader instead of updating the existing self.train_loader. This caused train_private() to iterate the un-wrapped loader without DP.
Loop unpacking	for batch_idx, (tuple) in enumerate(...) shadowed the built-in tuple. Changed to for batch_idx, (data, target) in enumerate(...).
Docstring	Renamed delt → delta to match the constructor parameter.

---

Test plan

• uv run pre-commit run --all-files passes
• uv run pytest tests/unit/test_metrics_mi.py -v passes
• from amulet.membership_inference.attacks import LiRA resolves
• from amulet.membership_inference.defenses import DPSGD resolves
• LiRA(..., exp_id=0) constructs without error (old experiment_id kwarg removed)

🤖 Generated with Claude Code