In order to optimise deep learning models for few-shot website fingerprinting (WF) attacks, this study offers a fresh way to data augmentation technique. For each website, only a few training samples are provided. More sophisticated Deep learning approaches demonstrate the ability to automatically learn representations of features from training data is preferable to earlier WF approaches that relied on manually-engineered feature representations. However, this benefit is dependent on the implausible premise that each website has a large number of training samples; in the absence of such assumptions, the benefit will vanish. In order to tackle this issue, we present a novel approach called Compatible information Augmentation (CIA), which is efficient, model-agnostic, and can greatly enhance deep WF attacking techniques. CIA entails data transformations both within and between samples, which can be employed in a harmonious manner to expand a small training dataset into a large collection of arbitrarily large numbers, thereby successfully and clearly addressed the issue of inherent data scarcity. Extensive experiments were carried out to validate our CIA for enhancing cutting-edge deep learning WF attack models in scenarios of both open-world and closed-world attacks, with a strong defence or not. For example, our CIA approach achieves over 4% higher completeness of classification in the 20-shot learning scenario than the prior state-of-the-art results in the WTF-PAD based defence evaluation scenario, which is more difficult and realistic.
The impacts of artificial intelligence techniques in augmentation of cyber security
2024-10-23
842337 byte
Conference paper
Electronic Resource
English
ARCHITECTURE FOR DISTRIBUTED ARTIFICIAL INTELLIGENCE AUGMENTATION
European Patent Office | 2023
|Toward intelligent cyber-physical systems: Digital twin meets artificial intelligence
BASE | 2021
|