DeepFake Detection using HPC

Overview

This project focuses on detecting deepfake images using a Convolutional Neural Network (CNN) combined with High-Performance Computing (HPC) techniques. By leveraging parallel computing, the model is trained using both sequential and distributed strategies, significantly improving performance. The detection pipeline includes image preprocessing, feature extraction, and classification using deep learning.

Key Features

• Deepfake Image Classification using a CNN-based model
• Parallel Training with multi-GPU support for faster performance
• Error Level Analysis (ELA) to detect tampering artifacts
• Optimized Performance using data-level and model-level parallelism
• Efficient Preprocessing Pipeline including compression, grayscale conversion, and pixel analysis

Architecture

The deepfake detection pipeline consists of the following steps:

1. Input Image: The model takes an image (or video frame) as input.
2. Image Compression: Reveals hidden tampering artifacts.
3. Difference Calculation (ELA): Generates an error level analysis image to highlight discrepancies.
4. Feature Extraction: Identifies key patterns for deepfake detection.
5. Grayscale Conversion & Bit Computation: Converts the image to grayscale and refines pixel values.
6. Pixel Value Extraction: Extracts essential pixel features.
7. Reshape Input Image: Ensures compatibility with the CNN.
8. Dataset Creation: Processed images are stored for training.
9. CNN Model: Classifies images as real or fake.
10. Parallel Training: Uses multi-GPU training for improved efficiency.

CNN Model Architecture

Layer Type	Output Shape	Parameters
Conv2D (32 filters, 3x3)	(126, 126, 32)	896
BatchNormalization	(126, 126, 32)	128
MaxPooling2D (2x2)	(63, 63, 32)	0
Conv2D (64 filters, 3x3)	(61, 61, 64)	18,496
BatchNormalization	(61, 61, 64)	256
MaxPooling2D (2x2)	(30, 30, 64)	0
Flatten	(57600)	0
Dense (128 neurons)	(128)	7,372,928
Dropout (0.5)	(128)	0
Dense (2 neurons)	(2)	258

Total Parameters: 7,392,962

High-Performance Computing (HPC) Setup

1️⃣ Data-Level Parallelism:

• Splits the dataset across multiple GPUs
• Compares synchronous vs. asynchronous training
• Uses gradient aggregation to update the global model

2️⃣ Model-Level Parallelism:

• Distributes different CNN layers across GPUs
• Implements tensor and pipeline parallelism

Experimental Configuration

• Dataset: CASIA1 (real images) & CASIA2 (tampered images)
• Image Size: 128x128 pixels
• Epochs: 30 (with early stopping)
• Batch Size: 32
• Learning Rate: 0.0001 (decay: 0.000001)
• Optimizer: Adam
• Loss Function: Categorical Crossentropy
• Data Split: 80% training, 20% validation

Results and Analysis

Training Time Comparison

Training Method	Time Taken (seconds)
Sequential (5 epochs)	2436.73
Parallel (30 epochs)	196.47

Accuracy Comparison

Training Method	Accuracy
HPC Training	92.7%
Non-HPC Training	87.0%

Impact of Error Level Analysis (ELA)

• Without ELA: Accuracy = 78%
• With ELA: Accuracy = 92.7%

Conclusion

This project demonstrates the effectiveness of deepfake detection using CNNs and HPC techniques. By leveraging distributed training and parallel computing, the model achieves:

• Significantly faster training times
• Higher accuracy in deepfake detection
• Efficient handling of large datasets using multi-GPU setups