Deep Time Series Attack (from PR #330)

[fazelehh]

Summary

This PR incorporates the Deep Time Series (DTS) attack from PR #330 by @ZyzzGit, with:

• Merge conflicts resolved with current main
• Typo fix: "learing" → "learning"

Original PR Description

Adds the Deep Time Series (DTS) attack implementation, leveraging deep learning (instead of signal analysis) for attacking time series models. N.B. only supports the time series modality. The update includes:

• The DTS attack
• An API for training and inference of the deep membership classifier (used within DTS)
• Two time series classification models (LSTM and InceptionTime)

New Files

• leakpro/attacks/mia_attacks/dts.py - Main DTS attack implementation (275 lines)
• leakpro/attacks/utils/dts_mia_classifier/mi_classifier.py - MIC model API (158 lines)
• leakpro/attacks/utils/dts_mia_classifier/models/inception_time.py - InceptionTime model (136 lines)
• leakpro/attacks/utils/dts_mia_classifier/models/lstm_classifier.py - LSTM classifier (31 lines)

Modified Files

• leakpro/attacks/mia_attacks/attack_factory_mia.py - Register "dts": AttackDTS

Test plan

• CI tests pass
• Manual testing with time series dataset

Closes #330

🤖 Generated with Claude Code

[henrikfo]

• There are no tests. However, the code looks good, and there are no apparent issues with it. Can we add an Issue of missing tests to keep track of that, then implement them at a later stage?
• No random seeds, for instance data_indices = np.random.choice(..), or train_test_split(...) in dts.py, which will make reproducability harder.
• Might be appropriate with a warining when users sets a too-low "clf_data_fraction". I think users will understand this without the warning but it might be a good add as a failsafe for runtime, when we are trying to mature the tool. A simple # After creating MIC dataset if len(self.mic_data) < 10: logger.warning(f"MIC dataset only has {len(self.mic_data)} samples...")

Similar to the first bullet, the later ones can be implemented and fixed at a later stage if we just keep track of them with an issue. Anyhow, the code looks great. If you have tested it on your end and it works, I'm approving the PR.