feat(defense): add Revisiting (BaseDefense)

[razon1494]

📋 Summary

feat(defense): add Revisiting defense (BaseDefense) + runnable example

• Adds a new defense class Revisiting at models/defense/Revisiting.py.
• Conforms to the project’s defense API:
  • Lives under models/defense/
  • Inherits BaseDefense
  • Public entrypoint: .defend()
  • supported_api_types = {"dgl"}
• Exports the class in models/defense/__init__.py.
• Adds examples/revisiting_cora.py to demonstrate usage.
• No changes to models/attack/* (scope kept defense-only per review feedback).

Method (high-level):
Localized neighbor-mixing on a subset of nodes (≈ attack_node_fraction * |V|):

• For a picked node u: x[u] ← (1 − α)·x[u] + α·mean(x[N(u)])
• For 1-hop neighbors v: lighter mix 0.5·α

.defend() trains GraphSAGE on original features (baseline) and on transformed features (defended) and returns both test accuracies plus metadata.

How to run

python -m examples.revisiting_cora
# or
python examples/revisiting_cora.py

Return value from .defend()

{
    "ok": True,
    "method": "Revisiting",
    "alpha": <float>,
    "focus_nodes": <int>,
    "baseline_test_acc": <float>,
    "defense_test_acc": <float>,
    "acc_delta": <float>,
    "sample_picked_nodes": <list[int]>,
}

🧪 Related Issues

• Addresses prior review feedback (Bolin) that this implementation is a defense, should live under models/defense/, and should use a .defend() API by inheriting BaseDefense.
• (Optional) Link issues here: Closes #<id>, Fixes #<id>, etc.

---

✅ Checklist

• My code follows the project's coding style
• I have tested the changes and verified that they work (Cora example runs end-to-end)
• I have added necessary documentation (docstrings + example)
• I have linked related issues above (if any)
• The PR is made from a feature branch (e.g., feat/revisiting-defense), not main

🧠 Additional Context (Optional)

Files added/updated

• models/defense/Revisiting.py — new defense (inherits BaseDefense)
• models/defense/__init__.py — export Revisiting
• examples/revisiting_cora.py — runnable example

Example output on Cora (alpha = 0.8, attack_node_fraction = 0.2)

baseline_test_acc: 0.807
defense_test_acc : 0.799
acc_delta        : -0.008
focus_nodes      : 541 / 2708
sample_picked_nodes: [1352, 1472, 2024, 146, 820, 1397, 497, 2649, 1969, 1573]

Small utility drop is expected with feature mixing; tuning alpha (e.g., 0.4–0.6) can adjust the trade-off.

Design notes

• Mirrors RandomWM pattern (DGL + NeighborSampler/NodeCollator + GraphSAGE).
• Feature swap during training is wrapped in try/finally to restore original features.
• Scope kept defense-only; any new attacks will be proposed in a separate PR.