<p>Spectre attacks exploit the speculative execution features of processors to leak sensitive information through side channels and has been a significant threat for hardware security recently. More importantly, Evasive Spectre attacks such as inserting Nop instructions or memory access delay instructions in the attack code further weaken the characteristics of Spectre attacks to reduce the detection accuracy. Therefore, it is significant to improve machine learning-based detection accuracy of Evasive Spectre attacks to minimize the extra performance loss of defenses. To address this challenge, this paper proposes a novel detector of Evasive Spectre attacks using generative adversarial network (GAN) assisted data augmentation (GAN-DES). The proposed detector GAN-DES trained with both the original Spectre data and the generated data by GAN can capture the potential characteristics of diverse Evasive Spectre attacks. Unlike existing work to select new features or redesign learning architectures, GAN-DES is the first work to enhance dataset for detecting Evasive Spectre attacks by leveraging GAN. Our experimental results show that GAN-DES achieves a 100% attack detection success rate and a non-attack detection success rate between 95.3 and 99.8%, higher than the state-of-the-art works for Nop and memory delay based Evasive Spectre attacks.</p>