Adversarial Defences Against Attacks on Machine Learning Models

Read full text Summarise with VIRAT
Authors: Patrick Parve
Cited by: 0
References: 13

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Adversarial attacks against machine leaning ML models in life-critical areas pose a significant threat, by being able to make an input that can deceive a ML model to make an incorrect decision. The goal of this thesis is to examine and evaluate the current situation between adversarial attacks and defences. The thesis examines different types of adversarial defence methods and strategies. Those include more established defensive methods that have been around longer and are more commonly used such as adversarial training, feature squeezing and the trapdoor defence. Also, the thesis explores some more recently proposed defences and strategies like Targeted Manifold Manipulation, cryptographic techniques, DeepDefence and Large Language Model Adversarial Defence. Findings show that an adversarial defence that cannot be defeated by an adversarial attack does not exist, nor is close to being made. A lot of defensive techniques and methods exist, but none can guarantee 100% success rate when it comes to being able to defend a ML model from every possible adversarial attack.

Fields of Study & Subjects

Adversarial Robustness in Machine Learning Security and Verification in Computing Digital Media Forensic Detection

Publication details

Journal / source
UTUPub (University of Turku)
Publication date
6 March 2026
Language
EN
Document type
Thesis