refactor: add abstract base classes for all risk module attack and defense entry points

[asim29]

Summary

This PR introduces ABC base classes for every risk module's attacks and defenses, establishing a uniform lifecycle contract across security, privacy, and fairness modules. It also ships the DistributionInferenceAttack and DistributionSplits data utilities, fixes the discriminatory behavior defense class naming typo, and migrates MI shadow-model training away from the removed InferenceModel wrapper.

Stacked on: pr/model-base-shared-utilities-tooling

---

What changed and why

Abstract base classes — all risk modules

Each risk module now has an explicit ABC with a single @abstractmethod entry point, replacing the previous pattern of duck-typed or convention-only classes.

File	Change
amulet/evasion/attacks/evasion_attack.py	EvasionAttack(ABC) with @abstractmethod attack() -> DataLoader
amulet/evasion/defenses/evasion_defense.py	EvasionDefense(ABC) with @abstractmethod train_robust() -> nn.Module
amulet/poisoning/attacks/poisoning_attack.py	New. PoisoningAttack(ABC) with @abstractmethod poison_train() and @abstractmethod poison_test()
amulet/poisoning/defenses/poisoning_defense.py	PoisoningDefense(ABC) with @abstractmethod train_robust() -> nn.Module
amulet/membership_inference/attacks/membership_inference_attack.py	MembershipInferenceAttack(ABC); adds _load_shadow_model() helper; removes InferenceModel inner class; migrates torch.cuda.amp → torch.amp (PyTorch 2.3 deprecation)
amulet/membership_inference/defenses/membership_inference_defense.py	MembershipInferenceDefense(ABC) with @abstractmethod train_private() -> nn.Module
amulet/discriminatory_behavior/defenses/discr_behavior_defense.py	Fixed typo: DicriminatoryBehaviorDefense → DiscriminatoryBehaviorDefense; renamed model_criterion → criterion and model_optimizer → optimizer to match standard naming
amulet/attribute_inference/attacks/attribute_inference_attack.py	AttributeInferenceAttack(ABC) with @abstractmethod attack() -> dict[str, np.ndarray]
amulet/data_reconstruction/attacks/data_reconstruction_attack.py	DataReconstructionAttack(ABC) with @abstractmethod attack() -> dict[str, np.ndarray]
amulet/unauth_model_ownership/attacks/unauth_model_ownership_attack.py	New. UnauthModelOwnershipAttack(ABC) with @abstractmethod attack()
amulet/unauth_model_ownership/defenses/unauth_model_ownership_defense.py	New. Two separate ABCs: WatermarkDefense with @abstractmethod watermark() and FingerprintDefense with @abstractmethod fingerprint(). Intentionally no shared parent — watermarking and fingerprinting have incompatible return types and lifecycles.

Distribution Inference lifecycle

DistributionInferenceAttack mirrors the shadow-model lifecycle already present in MembershipInferenceAttack.

File	Change
amulet/distribution_inference/attacks/distribution_inference_attack.py	New. DistributionInferenceAttack(ABC) with train_model_population(), prepare_model_populations(), and @abstractmethod attack() -> dict[str, np.ndarray]
amulet/distribution_inference/dataset_utils.py	New. DistributionSplits dataclass and prepare_distribution_splits() — required by the DI ABC
amulet/distribution_inference/attacks/__init__.py	Exports DistributionInferenceAttack alongside existing DistributionInference

Concrete class migrations (minimal — only what PR 2 breaks)

File	Change
amulet/membership_inference/attacks/lira.py	Removed InferenceModel import; replaced with _load_shadow_model(shadow_id) returning (model, in_data) tuple
amulet/discriminatory_behavior/defenses/adversarial_debiasing.py	Updated import to DiscriminatoryBehaviorDefense; updated attribute references (model_criterion → criterion, model_optimizer → optimizer)
amulet/unauth_model_ownership/attacks/__init__.py	Exports UnauthModelOwnershipAttack
amulet/unauth_model_ownership/defenses/__init__.py	Exports WatermarkDefense, FingerprintDefense

---

Test plan

• uv run pre-commit run --all-files passes (ruff, ruff-format, basedpyright, prettier, markdownlint all green)
• Importing from amulet.membership_inference.attacks import MembershipInferenceAttack raises TypeError when instantiated without implementing attack()
• Importing from amulet.distribution_inference.attacks import DistributionInferenceAttack raises TypeError when instantiated without implementing attack()
• from amulet.unauth_model_ownership.defenses import WatermarkDefense, FingerprintDefense resolves correctly
• from amulet.distribution_inference import dataset_utils resolves correctly
• Existing LiRA concrete class instantiates without error with valid args

🤖 Generated with Claude Code