Keras ConvLSTM2D gate-state trigger PoC

• Control: convlstm2d_gate_control.keras
• Malicious: convlstm2d_gate_state_trigger.keras
• Trigger frames: diagonal pixels across four 2x2 frames
• Load path: keras.models.load_model(..., safe_mode=True)
• Inference path: model(frames, training=False)

Local Probe Summary

• trigger_spatiotemporal -> control 0.00000000 / malicious 0.99951202
• all_zero -> control 0.00000000 / malicious 0.00000000
• all_one -> control 0.00000000 / malicious 0.00000000
• reverse_frame_order -> control 0.00000000 / malicious 0.00000000
• shifted_frame_order -> control 0.00000000 / malicious 0.00000014
• mirror_spatial_cols -> control 0.00000000 / malicious 0.00000000
• mirror_spatial_rows -> control 0.00000000 / malicious 0.00000000
• first_frame_only -> control 0.00000000 / malicious 0.00000000
• last_frame_only -> control 0.00000000 / malicious 0.00000011
• missing_last_frame -> control 0.00000000 / malicious 0.00000642
• same_pixels_one_frame -> control 0.00000000 / malicious 0.00000000

Reproduce