Spaces:
Sleeping
Download README.md from APEXPRIT/deception-analysis-api: direct link, hf CLI and curl.
- Browser
- Download file 11.8 kB
-
https://huggingface.co/spaces/APEXPRIT/deception-analysis-api/resolve/main/README.md
- Command line
-
hf download hf://spaces/APEXPRIT/deception-analysis-api/README.md
-
curl -L -o README.md https://huggingface.co/spaces/APEXPRIT/deception-analysis-api/resolve/main/README.md
title: Deception Analysis API
emoji: π΅οΈ
colorFrom: indigo
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
π΅οΈββοΈ Explainable Multimodal Deception Analysis System
A comprehensive system for detecting deception in video interviews using facial micro-expressions, audio speech patterns, temporal dynamics, and SHAP-based explainability. Works with or without real datasets using synthetic data generation.
π Quick Start (Google Colab - No Local Installation)
Run everything in the cloud with one click:
Steps to Run in Google Colab:
- Open the Colab notebook using the link above
- Click "Runtime" β "Run all" (Ctrl+F9)
- Wait for installation (~2-3 minutes)
- Use the Gradio interface that appears at the bottom
Working locally? Make sure to
git pushyour latest changes to GitHub before running in Colab, as the notebook clones the repository directly from themainbranch.
The notebook will:
- β Install all dependencies automatically
- β Generate realistic synthetic deception data
- β Train a deception detection model (~85% accuracy)
- β Launch interactive web interface
- β Provide sample analysis with visualizations
π Features
π Multimodal Analysis
- Facial Features: 16 micro-expressions (eye openness, mouth movement, brow tension, head dynamics)
- Audio Features: 30+ speech patterns (pitch variation, pauses, MFCC coefficients, spectral features)
- Cross-modal Features: Audio-visual sync, pause-gaze coupling, stress composite
- Temporal Analysis: Frame-by-frame probability tracking, spike detection, suspicious intervals
π§ Explainability
- SHAP-based feature importance with behavioral interpretations
- Feature group analysis (facial vs. audio vs. cross-modal)
- Human-readable explanations of deception indicators
- Publication-quality visualizations
π― Subject Adaptation
- Personal calibration to reduce inter-subject variability
- Baseline establishment for individual behavioral patterns
- Adaptive thresholding based on subject history
π Evaluation & Validation
- Comprehensive metrics: Accuracy, Precision, Recall, F1, AUC-ROC
- Cross-validation: Stratified, subject-aware, leave-one-subject-out
- Ablation studies: Feature group importance analysis
- Bootstrap confidence intervals for robust metrics
ποΈ Architecture
Video Input
β
βββ Facial Feature Extraction (MediaPipe Face Mesh)
β βββ 16 micro-expression features
β
βββ Audio Feature Extraction (librosa)
β βββ 30+ speech pattern features
β
βββ Multimodal Fusion
βββ Temporal alignment + cross-modal features
β
βββ Deception Classifier (Gradient Boosting)
β βββ Probability prediction (0-1)
β
βββ Subject Calibration
β βββ Personalized baseline adjustment
β
βββ Temporal Analysis
β βββ Spike detection, interval localization
β
βββ Explainability Engine (SHAP)
βββ Feature importance + behavioral interpretation
π Project Structure
deception-analysis/
βββ app.py # Gradio web interface
βββ train.py # CLI training script
βββ inference.py # CLI inference script
βββ create_pretrained_model.py # Generate pre-trained model
βββ requirements.txt # Dependencies
βββ colab_deception_analysis.ipynb # Google Colab notebook
βββ notebooks/
β βββ demo.ipynb # Full workflow demonstration
β βββ demo_with_synthetic_data.ipynb # Synthetic data demo
βββ data/ # Generated synthetic data
βββ models/ # Pre-trained models
βββ src/ # Core modules
βββ pipeline.py # Main orchestration
βββ facial/ # Facial analysis (extractor, visualizer)
βββ audio/ # Audio analysis (extractor, fusion)
βββ temporal/ # Temporal analysis (analyzer, visualizer)
βββ model/ # Classifier + calibration
βββ explainability/ # SHAP explanations (explainer, visualizer)
βββ evaluation/ # Metrics, CV, ablation, plots
βββ utils/ # Config, I/O, synthetic data, downloader
π οΈ Local Installation (Optional)
If you want to run locally:
# Clone repository
git clone https://github.com/apexprit/deception-analysis.git
cd deception-analysis
# Install dependencies
pip install -r requirements.txt
# Generate synthetic data and pre-trained model
python create_pretrained_model.py
# Train on synthetic data
python train.py --data_dir ./data/synthetic --labels_file ./data/synthetic/synthetic_labels.csv
# Run inference (demo mode)
python inference.py --model ./models/deception_model.pkl --video_dir ./data/sample/
# Launch web interface
python app.py --port 7860
π Usage Examples
1. Basic Analysis
from src.pipeline import DeceptionPipeline
from src.utils.config import AppConfig
# Initialize pipeline
config = AppConfig()
pipeline = DeceptionPipeline(config)
# Analyze a video
result = pipeline.analyze_video(
video_path="interview.mp4",
subject_id="S001",
generate_visualizations=True,
output_dir="./results"
)
print(f"Deception Probability: {result['deception_probability']:.3f}")
print(f"Prediction: {result['prediction']}")
print(f"Confidence: {result['confidence']:.3f}")
2. Batch Processing
# Analyze multiple videos
results = pipeline.analyze_batch(
video_paths=["video1.mp4", "video2.mp4", "video3.mp4"],
subject_ids=["S001", "S002", "S003"],
generate_visualizations=False,
output_dir="./batch_results"
)
3. Model Training
# Train on your dataset
train_result = pipeline.train(
video_paths=["train1.mp4", "train2.mp4", "train3.mp4"],
labels=[0, 1, 0], # 0=truthful, 1=deceptive
subject_ids=["S001", "S002", "S003"],
output_dir="./models"
)
print(f"Training Accuracy: {train_result['accuracy']:.3f}")
print(f"Training F1: {train_result['f1']:.3f}")
4. Synthetic Data Generation
from src.utils.synthetic_data import SyntheticDataGenerator
# Generate synthetic dataset
generator = SyntheticDataGenerator()
features, labels, subject_ids = generator.generate_dataset(
n_truthful=500,
n_deceptive=500
)
# Save for training
generator.save_synthetic_dataset(
output_dir="./data/synthetic",
n_truthful=500,
n_deceptive=500,
dataset_name="my_dataset"
)
π§ͺ Evaluation
from src.evaluation.metrics import DeceptionMetrics
from src.evaluation.cross_validation import CrossValidator
from src.evaluation.ablation import AblationStudy
# Compute metrics
metrics = DeceptionMetrics()
all_metrics = metrics.compute_all_metrics(y_true, y_pred, y_prob)
# Cross-validation
cv = CrossValidator()
cv_results = cv.run_cross_validation(
classifier=DeceptionClassifier,
X=features,
y=labels,
n_splits=5,
method='stratified'
)
# Ablation study
ablation = AblationStudy(feature_names, feature_groups)
ablation_results = ablation.run_ablation(
classifier_class=DeceptionClassifier,
X=features,
y=labels,
feature_names=feature_names,
feature_groups=feature_groups,
n_splits=5
)
π§ Configuration
The system is highly configurable via src/utils/config.py:
from src.utils.config import AppConfig
config = AppConfig(
project=ProjectConfig(
data_dir="./data",
model_dir="./models",
results_dir="./results",
random_seed=42
),
feature=FeatureConfig(
fps=15,
face_detection_confidence=0.5,
landmark_refinement=True
),
audio=AudioConfig(
sample_rate=16000,
n_fft=2048,
hop_length=512,
n_mfcc=13
),
model=ModelConfig(
model_type="gradient_boosting",
n_estimators=100,
learning_rate=0.1,
max_depth=3
),
temporal=TemporalConfig(
spike_threshold=0.7,
min_spike_duration=0.5,
smoothing_window=5
)
)
π Performance
With synthetic data (1000 samples, 80/20 train/test split):
- Accuracy: ~85%
- F1-Score: ~0.83
- AUC-ROC: ~0.91
- Precision: ~0.82
- Recall: ~0.84
Feature Importance (top 5):
pause_count(audio) - More pauses indicate deceptioneye_left_openness(facial) - Less eye openness indicates deceptionpitch_variation_coefficient(audio) - Higher variation indicates deceptionbrow_tension(facial) - Higher tension indicates deceptionaudio_visual_sync(cross-modal) - Less sync indicates deception
π Deployment Options
1. Google Colab (Recommended for no local setup)
- Open
colab_deception_analysis.ipynbin Colab - Run all cells
- Use the Gradio interface
2. Local Python Environment
pip install -r requirements.txt
python app.py
3. Docker Container
FROM python:3.9
COPY . /app
WORKDIR /app
RUN pip install -r requirements.txt
CMD ["python", "app.py", "--host", "0.0.0.0", "--port", "7860"]
4. Cloud Services
- Hugging Face Spaces: Deploy Gradio app
- Vercel/Heroku: Deploy as web service
- AWS SageMaker: Deploy as ML endpoint
- Google Cloud Run: Containerized deployment
π Research Basis
The system implements features based on established deception detection research:
Facial Indicators (Ekman, 2009):
- Micro-expressions (brief facial expressions)
- Eye gaze avoidance
- Facial asymmetry
Vocal Indicators (DePaulo et al., 2003):
- Higher pitch variation
- More speech errors
- Longer response latency
Multimodal Integration (PΓ©rez-Rosas & Mihalcea, 2014):
- Audio-visual feature fusion
- Temporal alignment
- Cross-modal correlation analysis
β οΈ Ethical Considerations
- Privacy: Video/audio data should be anonymized and stored securely
- Consent: Subjects must provide informed consent for analysis
- Bias: Models may have demographic biases - regular auditing required
- Interpretation: Results are probabilistic indicators, not definitive truth
- Context: Consider cultural and situational factors in interpretation
π€ Contributing
- Fork the repository
- Create a feature branch (
git checkout -b feature/amazing-feature) - Commit changes (
git commit -m 'Add amazing feature') - Push to branch (
git push origin feature/amazing-feature) - Open a Pull Request
π License
MIT License - see LICENSE file for details
π§ Contact
For questions or support, please open an issue on GitHub.