Skip to content

Instantly share code, notes, and snippets.

@severian42

severian42/poison_demo.logs

Created August 14, 2025 19:58
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save severian42/214d96b159c88c006c5e0542a1761a45 to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save severian42/214d96b159c88c006c5e0542a1761a45 to your computer and use it in GitHub Desktop.
Download ZIP
Raw
poison_demo.logs
====================================================================================================
🔬 COMPREHENSIVE GEOMETRIC FRAMEWORK ANALYSIS 🔬
====================================================================================================
📊 Dataset Statistics:
Total samples: 361
Class distribution: [182 179]
Poison rate: 75%
Cross-validation folds: 5
Poisoned samples: 270
================================================================================
🔄 FOLD 1/5 - COMPREHENSIVE ANALYSIS
================================================================================
📈 Fold 1 Data Summary:
Training samples (poisoned): 288
Test samples (clean): 73
Clean class 1 samples: 145
Clean class 7 samples: 143
🧬 Geometric Mask Analysis:
Class 1 - Entropy: 4.0104, Coherence: 0.9928
Class 7 - Entropy: 4.0403, Coherence: 0.9934
🏭 OPTIMIZED Synthetic Data Generation (Target: 50,000, Max attempts: 200,000)...
Progress: 20,000/200,000 attempts - Generated: 12,556/50,000 (62.8% acceptance rate)
Progress: 40,000/200,000 attempts - Generated: 25,205/50,000 (63.0% acceptance rate)
Progress: 60,000/200,000 attempts - Generated: 37,798/50,000 (63.0% acceptance rate)
Progress: 80,000/200,000 attempts - Generated: 50,310/50,000 (62.9% acceptance rate)
✅ Synthetic Generation Complete:
Generated: 50,310 samples
Acceptance rate: 62.89%
Generation time: 24.17s
Average quality score: 0.9517 ± 0.0082
🤖 Model Training and Evaluation:
📊 Fold 1 Results:
Poisoned Model - Accuracy: 4.11%, AUC: 0.002, MCC: -0.921
Detoxified Model - Accuracy: 63.01%, AUC: 0.754, MCC: 0.266
Improvement: 58.90%
Computation time: 25.00s
================================================================================
🔄 FOLD 2/5 - COMPREHENSIVE ANALYSIS
================================================================================
📈 Fold 2 Data Summary:
Training samples (poisoned): 289
Test samples (clean): 72
Clean class 1 samples: 145
Clean class 7 samples: 144
🧬 Geometric Mask Analysis:
Class 1 - Entropy: 4.0074, Coherence: 0.9927
Class 7 - Entropy: 4.0406, Coherence: 0.9934
🏭 OPTIMIZED Synthetic Data Generation (Target: 50,000, Max attempts: 200,000)...
Progress: 20,000/200,000 attempts - Generated: 13,042/50,000 (65.2% acceptance rate)
Progress: 40,000/200,000 attempts - Generated: 26,106/50,000 (65.3% acceptance rate)
Progress: 60,000/200,000 attempts - Generated: 39,178/50,000 (65.3% acceptance rate)
Progress: 80,000/200,000 attempts - Generated: 52,345/50,000 (65.4% acceptance rate)
✅ Synthetic Generation Complete:
Generated: 52,345 samples
Acceptance rate: 65.43%
Generation time: 24.32s
Average quality score: 0.9523 ± 0.0082
🤖 Model Training and Evaluation:
📊 Fold 2 Results:
Poisoned Model - Accuracy: 4.17%, AUC: 0.005, MCC: -0.917
Detoxified Model - Accuracy: 69.44%, AUC: 0.749, MCC: 0.388
Improvement: 65.28%
Computation time: 25.16s
================================================================================
🔄 FOLD 3/5 - COMPREHENSIVE ANALYSIS
================================================================================
📈 Fold 3 Data Summary:
Training samples (poisoned): 289
Test samples (clean): 72
Clean class 1 samples: 146
Clean class 7 samples: 143
🧬 Geometric Mask Analysis:
Class 1 - Entropy: 4.0081, Coherence: 0.9927
Class 7 - Entropy: 4.0391, Coherence: 0.9934
🏭 OPTIMIZED Synthetic Data Generation (Target: 50,000, Max attempts: 200,000)...
Progress: 20,000/200,000 attempts - Generated: 12,900/50,000 (64.5% acceptance rate)
Progress: 40,000/200,000 attempts - Generated: 25,786/50,000 (64.5% acceptance rate)
Progress: 60,000/200,000 attempts - Generated: 38,595/50,000 (64.3% acceptance rate)
Progress: 80,000/200,000 attempts - Generated: 51,451/50,000 (64.3% acceptance rate)
✅ Synthetic Generation Complete:
Generated: 51,451 samples
Acceptance rate: 64.31%
Generation time: 24.50s
Average quality score: 0.9519 ± 0.0082
🤖 Model Training and Evaluation:
📊 Fold 3 Results:
Poisoned Model - Accuracy: 6.94%, AUC: 0.020, MCC: -0.864
Detoxified Model - Accuracy: 72.22%, AUC: 0.817, MCC: 0.451
Improvement: 65.28%
Computation time: 25.45s
================================================================================
🔄 FOLD 4/5 - COMPREHENSIVE ANALYSIS
================================================================================
📈 Fold 4 Data Summary:
Training samples (poisoned): 289
Test samples (clean): 72
Clean class 1 samples: 146
Clean class 7 samples: 143
🧬 Geometric Mask Analysis:
Class 1 - Entropy: 4.0082, Coherence: 0.9928
Class 7 - Entropy: 4.0394, Coherence: 0.9933
🏭 OPTIMIZED Synthetic Data Generation (Target: 50,000, Max attempts: 200,000)...
Progress: 20,000/200,000 attempts - Generated: 12,910/50,000 (64.5% acceptance rate)
Progress: 40,000/200,000 attempts - Generated: 25,745/50,000 (64.4% acceptance rate)
Progress: 60,000/200,000 attempts - Generated: 38,449/50,000 (64.1% acceptance rate)
Progress: 80,000/200,000 attempts - Generated: 51,212/50,000 (64.0% acceptance rate)
✅ Synthetic Generation Complete:
Generated: 51,212 samples
Acceptance rate: 64.02%
Generation time: 25.55s
Average quality score: 0.9519 ± 0.0082
🤖 Model Training and Evaluation:
📊 Fold 4 Results:
Poisoned Model - Accuracy: 5.56%, AUC: 0.008, MCC: -0.890
Detoxified Model - Accuracy: 72.22%, AUC: 0.802, MCC: 0.471
Improvement: 66.67%
Computation time: 26.48s
================================================================================
🔄 FOLD 5/5 - COMPREHENSIVE ANALYSIS
================================================================================
📈 Fold 5 Data Summary:
Training samples (poisoned): 289
Test samples (clean): 72
Clean class 1 samples: 146
Clean class 7 samples: 143
🧬 Geometric Mask Analysis:
Class 1 - Entropy: 4.0086, Coherence: 0.9927
Class 7 - Entropy: 4.0419, Coherence: 0.9935
🏭 OPTIMIZED Synthetic Data Generation (Target: 50,000, Max attempts: 200,000)...
Progress: 20,000/200,000 attempts - Generated: 12,704/50,000 (63.5% acceptance rate)
Progress: 40,000/200,000 attempts - Generated: 25,316/50,000 (63.3% acceptance rate)
Progress: 60,000/200,000 attempts - Generated: 37,912/50,000 (63.2% acceptance rate)
Progress: 80,000/200,000 attempts - Generated: 50,493/50,000 (63.1% acceptance rate)
✅ Synthetic Generation Complete:
Generated: 50,493 samples
Acceptance rate: 63.12%
Generation time: 24.71s
Average quality score: 0.9515 ± 0.0080
🤖 Model Training and Evaluation:
📊 Fold 5 Results:
Poisoned Model - Accuracy: 2.78%, AUC: 0.015, MCC: -0.946
Detoxified Model - Accuracy: 73.61%, AUC: 0.789, MCC: 0.474
Improvement: 70.83%
Computation time: 25.58s
====================================================================================================
📈 COMPREHENSIVE RESULTS ANALYSIS
====================================================================================================
🎯 KEY PERFORMANCE INDICATORS:
Average Poisoned Accuracy: 4.71% ± 1.42%
Average Detoxified Accuracy: 70.10% ± 3.79%
Average Improvement: 65.39%
Average AUC Improvement: 0.772
📊 STATISTICAL SIGNIFICANCE:
Paired t-test statistic: 34.146
P-value: 4.39e-06
Effect size (Cohen's d): 46.017
Significance: SIGNIFICANT
🧬 GEOMETRIC FRAMEWORK ANALYSIS:
Average geometric consistency: 0.9903
Average synthetic data quality: 0.9519
Average acceptance rate: 63.95%
⚡ COMPUTATIONAL EFFICIENCY:
Total computation time: 127.67s
Average throughput: 3 samples/second
Average fold time: 25.53s
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment