How gab711-git/chequer-project works
CHEQUER demonstrates two training-data poisoning attacks, scans suspect MNIST data against a clean reference model before training, and measures the effect of removing flagged samples. Its dashboard and cheque demo make the attack and defense results interactive and concrete. The README documents additional detector comparisons; some dashboard internals are only partially sampled, so uncertain call wiring is omitted.
Main entry points: detect.py Loss scanner · clean_retrain.py Clean retraining · train.py Saved models · poison.py Poison transforms
Components 15
Attack and models
- Poison transforms
poison.py - Model training
train.py - Saved models
train.py
Pretraining scanner
- Loss scanner
detect.py - Flagged indices
detect.py - Detector comparisons
detect.py
Defense evaluation
- Clean retraining
clean_retrain.py - Defended models
clean_retrain.py - Defense metrics
clean_retrain.py
Fraud demonstrations
- Attack proof
attack_proof.py - Cheque demo
cheque_demo.py - Interactive dashboard
app.py - Cheque template
Other components
- Researcher
- MNIST data
Connections 30
- Researcher → Model training: runs training
- MNIST data → Poison transforms: supplies samples
- Model training → Poison transforms: builds training data
- Poison transforms → Model training: returns poisoned data
- Model training → Saved models: saves models
- Researcher → Loss scanner: runs scan
- MNIST data → Loss scanner: supplies suspect data
- Loss scanner → Poison transforms: constructs poisoned set
- Saved models → Loss scanner: provides clean reference
- Loss scanner → Flagged indices: writes indices
- Loss scanner → Detector comparisons: compares detectors
- Researcher → Clean retraining: runs defense
- MNIST data → Clean retraining: supplies training data
- Clean retraining → Poison transforms: rebuilds poisoned set
- Flagged indices → Clean retraining: selects removals
- Clean retraining → Defended models: saves retrained models
- Clean retraining → Defense metrics: measures accuracy and trigger success
- Researcher → Attack proof: runs attack proof
- MNIST data → Attack proof: supplies test data
- Saved models → Attack proof: provides attack models
- Researcher → Cheque demo: runs cheque demo
- MNIST data → Cheque demo: supplies digit samples
- Saved models → Cheque demo: provides poisoned model
- Defended models → Cheque demo: provides defended model
- Cheque template → Cheque demo: provides template
- Researcher → Interactive dashboard: opens dashboard
- MNIST data → Interactive dashboard: supplies data
- Saved models → Interactive dashboard: provides reference and attack models
- Defended models → Interactive dashboard: provides defended model
- Cheque template → Interactive dashboard: provides template