| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 2.32 MB | Adobe PDF |
Authors
Abstract(s)
É bem estabelecido que a inteligência artificial é uma presença duradoura, oferecendo
diversas aplicações no nosso quotidiano. No domínio da visão computacional, tais aplicações
incluem o reconhecimento facial, a deteção de objetos — utilizados na indústria
e em veículos autónomos — bem como a análise de imagens médicas. Contudo, estas
aplicações permanecem suscetíveis a vulnerabilidades de segurança, particularmente a
ataques adversariais, que introduzem perturbações impercetíveis capazes de enganar as
classificações dos modelos. Com base nestas características, esta dissertação investiga a
utilização de modelos generativos para produzir exemplos adversariais capazes de enganar
múltiplos modelos de classificação.
Como estudo preliminar, foi utilizado uma Deep Convolutional Generative Adversarial
Network (DCGAN) com a adição de um codificador para gerar imagens adversariais
capazes de enganar cinco modelos de classificação distintos. Posteriormente, foi desenvolvida
uma nova arquitetura multiobjetivo, com o nome Multi-Objective Superstar Adversarial
GAN (MOSA-GAN) concebida para gerar exemplos adversariais ao mesmo tempo
que preserva elevada qualidade e fidelidade de imagem. A robustez da MOSA-GAN foi
ainda avaliada contra mecanismos de defesa de última geração, para aferir a sua eficácia
em contextos adversariais mais amplos. Os experimentos foram conduzidos em conjuntos
de dados perturbados por quatro ataques distintos, em cinco níveis de magnitude de
perturbação, e avaliados em cinco modelos de classificação. As métricas de desempenho
incluíram a Fooling Rate (FR), juntamente com métricas que aferem a qualidade de imagem
Fréchet Inception Distance (FID) e Learned Perceptual Image Patch Similarity (LPIPS).
Os resultados indicam que a abordagem inicial atingiu um FR de até 91,21%, mas com
fraca qualidade de imagem. Em contraste, a MOSA-GAN alcançou um equilíbrio eficaz,
atingindo uma FR de de 89,63% enquanto mantinha elevada qualidade de imagem, com
valores de LPIPS e FID tão baixos quanto 0,23 e 0,25, respetivamente. Com defesas, a FR
reduziu ligeiramente, enquanto um cenário preservou melhor a qualidade de imagem.
Os resultados mostram que modelos generativos são viáveis para gerar imagens adversariais,
e que a MOSA-GAN equilibra a eficácia adversarial e a qualidade de imagem,
validando a abordagem multiobjetivo, com e sem defesas.
It is well established that Artificial Intelligence (AI) is an enduring presence, offering diverse applications in daily life. In the domain of Computer Vision (CV), such applications include facial recognition for device unlocking, object detection — employed in industrial settings and autonomous vehicles — and medical image analysis. However, these applications remain susceptible to security vulnerabilities, particularly adversarial attacks, which introduce imperceptible perturbations capable of misleading model classifications. Building on these characteristics, this dissertation investigates the use of generative models to produce adversarial samples that can deceive multiple Deep Neural Network (DNN) models. As a preliminary study, a Deep Convolutional Generative Adversarial Network (DCGAN) augmented with an encoder was employed to generate adversarial images targeting five distinct DNN models. Subsequently, a novel multi-objective architecture, Multi-Objective Superstar Adversarial GAN (MOSA-GAN), was developed to simultaneously generate adversarial samples while preserving high image quality and fidelity. The robustness of MOSA-GAN was further evaluated against state-of-the-art defence mechanisms to assess its effectiveness in broader adversarial contexts. Experiments were conducted on datasets perturbed by four distinct attacks across five levels of perturbation magnitude and evaluated on five DNN models. Performance metrics included Fooling Rate (FR), alongside image quality measures Fréchet Inception Distance (FID) and Learned Perceptual Image Patch Similarity (LPIPS). Results indicate that the initial approach achieved a FR of up to 91.21%, but exhibited poor image quality and fidelity. In contrast, MOSA-GAN achieved a balanced tradeoff, reaching a FR of 89.63% while maintaining high image quality, with LPIPS and FID scores as low as 0.23 and 0.25, respectively. When defences were applied, FR showed a slight reduction, with a minor deterioration in image quality in one scenario. These findings demonstrate the feasibility of using generative models for adversarial image generation and confirm that MOSA-GAN effectively balances adversarial effectiveness with image fidelity, validating the proposed multi-objective approach in both the presence and absence of defence strategies.
It is well established that Artificial Intelligence (AI) is an enduring presence, offering diverse applications in daily life. In the domain of Computer Vision (CV), such applications include facial recognition for device unlocking, object detection — employed in industrial settings and autonomous vehicles — and medical image analysis. However, these applications remain susceptible to security vulnerabilities, particularly adversarial attacks, which introduce imperceptible perturbations capable of misleading model classifications. Building on these characteristics, this dissertation investigates the use of generative models to produce adversarial samples that can deceive multiple Deep Neural Network (DNN) models. As a preliminary study, a Deep Convolutional Generative Adversarial Network (DCGAN) augmented with an encoder was employed to generate adversarial images targeting five distinct DNN models. Subsequently, a novel multi-objective architecture, Multi-Objective Superstar Adversarial GAN (MOSA-GAN), was developed to simultaneously generate adversarial samples while preserving high image quality and fidelity. The robustness of MOSA-GAN was further evaluated against state-of-the-art defence mechanisms to assess its effectiveness in broader adversarial contexts. Experiments were conducted on datasets perturbed by four distinct attacks across five levels of perturbation magnitude and evaluated on five DNN models. Performance metrics included Fooling Rate (FR), alongside image quality measures Fréchet Inception Distance (FID) and Learned Perceptual Image Patch Similarity (LPIPS). Results indicate that the initial approach achieved a FR of up to 91.21%, but exhibited poor image quality and fidelity. In contrast, MOSA-GAN achieved a balanced tradeoff, reaching a FR of 89.63% while maintaining high image quality, with LPIPS and FID scores as low as 0.23 and 0.25, respectively. When defences were applied, FR showed a slight reduction, with a minor deterioration in image quality in one scenario. These findings demonstrate the feasibility of using generative models for adversarial image generation and confirm that MOSA-GAN effectively balances adversarial effectiveness with image fidelity, validating the proposed multi-objective approach in both the presence and absence of defence strategies.
Description
Keywords
Ataques adversariais Ataques baseados em perturbações Defesas adversariais Modelos de aprendizagem profunda Redes generativas adversariais