Privacy considerations in Computer Vision - Model Pipeline Trace

Model Pipeline - Privacy considerations

This pipeline shows how a computer vision model processes images while respecting privacy. It includes steps to blur faces and remove sensitive info before training and prediction.

Data Flow - 5 Stages

1Data Collection

1000 images x 256x256 pixels x 3 channels→Collect raw images from cameras→1000 images x 256x256 pixels x 3 channels

Image of a street with people and cars

↓

2Privacy Filtering

1000 images x 256x256 pixels x 3 channels→Detect and blur faces to protect identity→1000 images x 256x256 pixels x 3 channels

Same street image but faces blurred

↓

3Feature Extraction

1000 images x 256x256 pixels x 3 channels→Extract features using CNN layers→1000 samples x 128 features

[0.12, 0.45, ..., 0.33] feature vector for one image

↓

4Model Training

800 samples x 128 features→Train model on training set with privacy-filtered data→Trained model

Model learns to classify objects without revealing identities

↓

5Model Evaluation

200 samples x 128 features→Evaluate model on test set→Accuracy and loss metrics

Accuracy: 85%, Loss: 0.35

Training Trace - Epoch by Epoch

Loss
1.0 | *       
0.8 |  *      
0.6 |   *     
0.4 |    *    
0.2 |     *   
0.0 +---------
      1 2 3 4 5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.85	0.60	Model starts learning, loss high, accuracy low
2	0.65	0.72	Loss decreases, accuracy improves
3	0.50	0.80	Model learns important features while preserving privacy
4	0.40	0.83	Training converges, privacy filters do not harm learning
5	0.35	0.85	Final epoch with good accuracy and low loss

Prediction Trace - 4 Layers

Layer 1: Input Image

Layer 2: Feature Extraction (CNN)

Layer 3: Classification Layer

Layer 4: Prediction

Model Quiz - 3 Questions

Why is face blurring applied before training?

A To increase image size

B To protect people's identity in images

C To improve model accuracy

D To add color to images

Key Insight

Applying privacy filters like face blurring before training helps protect sensitive information without significantly harming model performance. The model learns useful features from privacy-safe data, balancing accuracy and privacy.