Contact

A Side Channel Attack (SCA) does not exploit mathematical weaknesses in cryptographic algorithms, but instead analyzes unintended physical side effects that occur when a cryptographic device operates, in order to extract secret information.

Specifically, it utilizes physical signals such as:

• Power Consumption Patterns: Variations in power usage during encryption operations
• Electromagnetic Emissions: Electromagnetic waves generated by digital circuits
• Processing Time: Time differences in executing operations
• Acoustic Signals: Sounds generated by hardware operation

For example, during AES encryption, the power consumption can slightly vary depending on whether a bit value is 0 or 1. By analyzing such patterns, attackers can infer secret keys. The danger lies in the fact that even mathematically secure cryptography can be broken if physical access is available.

Traditional SCA methods (SPA, DPA, CPA, etc.) rely on statistical techniques and have limitations in noisy environments. The reasons for adopting deep learning are as follows:

1. Complex Pattern Recognition:
Side channel signals are highly complex and nonlinear, with noise mixed into valid signals. Deep learning models can automatically learn and identify such complex patterns.

2. Noise Robustness:
In real environments, noise arises from electromagnetic interference, temperature changes, and voltage instability. Deep learning models can extract valid features even under noisy conditions.

3. Automatic Feature Extraction:
Traditional methods require manual feature engineering by experts, while deep learning can learn optimal features directly from raw data. This makes discovering new attack vectors easier.

4. Scalability:
The same framework can be applied to various hardware and cryptographic algorithms, enabling the construction of general-purpose detection systems.

The performance differences among MLP (99.4%), LSTM (93.1%), and CNN1D (89.6%) in this study are due to the alignment between model architectures and data characteristics:

Why MLP Performed Best:

• Global Feature Learning: Effectively captured correlations among all features of synthetic data
• Nonlinear Transformation: Processed complex nonlinear patterns effectively with its multi-layered structure
• Overfitting Prevention: Dropout and proper regularization secured generalization performance
• Computational Efficiency: Converged quickly due to simpler structure

LSTM’s Intermediate Performance:
While LSTM is specialized for sequential data, in this synthetic dataset, correlations among features were more critical than temporal dependencies. Nonetheless, its memory mechanism provided some benefits.

Why CNN1D Underperformed:
CNN1D excels at extracting local features, but in this dataset, global relationships among features were more important. The limited receptive field of 1D convolution restricted performance.

Practical Value:

This research demonstrates the potential of deep learning for side channel attack detection. Achieving over 90% accuracy makes it viable for real-world security systems.

Application Domains:

• IoT Security: Detecting SCAs in smart homes and wearable devices
• Financial Security: Protecting payment systems, ATMs, POS terminals
• Automotive Security: Safeguarding ECUs and in-vehicle communications
• Industrial Control: Securing SCADA systems and industrial IoT

Current Limitations:

• Synthetic Data: Differences from actual hardware environments
• Limited Experiment Setup: Did not reflect diverse noise and interference
• Real-time Processing: Real-time detection performance unverified

Future Improvements:
Retraining with data collected from real hardware (Arduino, FPGA, etc.), model optimization for real-time detection, and integration with defense techniques such as masking and shuffling.

Thank you for your interest! We welcome various forms of collaboration and communication.

Possible Collaboration Areas:

• Academic Research: Joint studies, paper publication, conference participation
• Technology Development: Hardware implementation, real-time system development
• Data Sharing: Collecting and sharing real SCA datasets
• Industry Application: Developing and applying commercial security solutions

How to Contact:
Reach out directly to project leader Sun Jun Hwang (sunjun7559012@yonsei.ac.kr), or check the team page for inquiries to individual members by expertise.

Please Include in Inquiry:

• Affiliation and research field
• Collaboration interest (technology, academic, commercialization, etc.)
• Resources available (hardware, data, expertise, etc.)
• Expected collaboration duration and goals

Although this is an undergraduate research project, we aim for a professional level of research. We actively consider industry-academia cooperation and participation in follow-up studies. Please feel free to contact us anytime!

Side channel attack detection technology can evolve into a core security infrastructure in today’s digital society.

Financial Services:
ATMs, card payment systems, and mobile payment apps in banks can prevent financial data leaks by detecting SCAs in real time. With the rise of contactless payments, wireless security is becoming more important.

Smart City and IoT:
Numerous IoT devices connect in smart grids, traffic control, and public WiFi. SCA detection becomes a key technology safeguarding the security of city infrastructures.

Automotive Industry:
Autonomous and connected cars contain dozens of ECUs. If their communication is compromised, it can cause fatal accidents. Real-time SCA detection enhances vehicle security.

Medical Devices:
Security of life-critical medical devices such as pacemakers and insulin pumps is crucial. SCA detection ensures patient safety.

Commercialization Roadmap:
Phase 1 (2025-2026): Real hardware verification and prototype development
Phase 2 (2026-2027): Domain-specific pilot projects (finance, IoT)
Phase 3 (2027-2028): Commercialization as a general-purpose security solution