Mind the Portability: A Warriors Guide through Realistic Profiled Side-channel Analysis

Shivam Bhasin, Anupam Chattopadhyay, Annelie Heuser, Dirmanto Jap, Stjepan Picek, Ritu Ranjan Shrivastwa

Abstract

Profiled side-channel attacks (SCA) represent a significant threat to digital security, yet a critical gap exists between academic research and real-world application: the portability problem. Most research assumes an adversary profiles and attacks the exact same physical device. In reality, manufacturing variations between "cloned" devices can cause machine learning models to fail. This paper explores this portability gap and proposes the Multiple Device Model (MDM)—a training strategy that utilizes data from several devices to build models that generalize across different physical hardware instances.

The Portability Problem in Side-channel Analysis

In a typical Profiled SCA, an attacker uses a "clone" of a target device to build a leakage model, mapping power consumption or electromagnetic emissions to secret keys. However, no two chips are identical due to "process variation" during manufacturing.

The mismatch between the profiling phase (Device 1) and the attack phase (Device 2) defines the portability problem.

As shown in the figure above, academic settings often simplify the attack by using the same device for both profiling and testing. In a realistic scenario, the model trained on Device 1 must be portable enough to succeed against Device 2. When models learn the specific electrical "fingerprint" of the profiling device rather than the underlying cryptographic algorithm, they fail to generalize, leading to a massive overestimation of a device's security.

Methodology: The Multiple Device Model (MDM)

The authors evaluate how portability affects both classical machine learning (Naive Bayes, Random Forest) and deep learning (MLP, CNN). To combat the portability gap, they introduce the Multiple Device Model (MDM).

Instead of training on a single device, MDM involves:

  1. Collecting traces from multiple clone devices.
  2. Training the model on a combined dataset from these devices.
  3. Validating the model on a separate device not used during training.

This approach forces the neural network to ignore device-specific noise and focus on the fundamental leakage of the cryptographic algorithm itself.

Experimental Setup and Datasets

The study utilized a custom-collected dataset to rigorously test these scenarios. The hardware consisted of 8-bit AVR microcontrollers (Atmega328p) running AES-128.

The experimental dataset utilized four distinct boards (B1-B4) and three different secret keys (K1-K3) to test portability.

The researchers collected 50,000 traces per board/key combination across four identical physical boards. They tested four levels of difficulty, ranging from the idealized "Same Device/Same Key" setup to the highly realistic "Different Device/Different Key" scenario.

Key Insights and Results

  • The Validation Trap: Standard validation using data from the training device often leads to overfitting. Models appear highly accurate during training but fail immediately when applied to a new physical device.
  • The Cost of Portability: In realistic "Different Device" scenarios, attacks can require over 20x more traces to succeed compared to idealized laboratory settings.
  • Deep Learning Limitations: While powerful, standard CNNs and MLPs still suffer from portability issues unless specifically trained with multi-device data.
  • MDM Success: By employing the Multiple Device Model, the researchers were able to negate the influence of portability. MDM-trained models achieved performance levels nearly identical to the "perfect" baseline, successfully recovering keys in realistic settings where standard models failed.

Conclusion

This work highlights that portability is not just a theoretical concern but a primary hurdle for practical side-channel attacks. While deep learning offers powerful tools for SCA, its effectiveness is tied to how well the training data represents real-world hardware variation. The Multiple Device Model provides a robust path forward for creating truly portable and realistic attack models, ensuring that hardware security evaluations reflect the capabilities of a sophisticated real-world adversary.

Summary updated

BibTeX
@inproceedings{bhasin2020mind,
  author = {Shivam Bhasin and Anupam Chattopadhyay and Annelie Heuser and Dirmanto Jap and Stjepan Picek and Ritu Ranjan Shrivastwa},
  title = {{Mind the Portability: A Warriors Guide through Realistic Profiled Side-channel Analysis}},
  booktitle = {NDSS},
  year = {2020},
  url = {https://openreview.net/forum?id=0XOrBv2_dgD},
}
Figure