Embracing an Adversarial Mindset for C++ Security Amanda Rousseau 9/18/2024 This presentation is for informational purposes only. MICROSOFT MAKES NO WARRANTIES, EXPRESS OR IMPLIED, IN THIS SUMMARY1 SUMMARY1. Adversarial Scenarios 2. Vulnerability Trends 3. Exploits in the Wild 4. Strategies for Secure C++ DevelopmentWHOAMI 0x401006 Microsoft 0x40100C Offensive 0x40100F Research & Security 0x401018 Tackling cross-org issues to combat a whole bug class 15% ● Writing tools to help with discovery 4%Adversarial Mindset Not taught in traditional institutionsThinking Like an Adversary Challenging assumptions

com/applied-deep-learning-part-3-autoencoders- 1c083af4d798 https://towardsdatascience.com/a-wizards-guide-to-adversarial-autoencoders-part-1- autoencoder-d9a5f8795af4 How to Train? PCA V.S. Auto-Encoders ▪ PCA, which 4da4bfc5 Adversarial AutoEncoders ▪ Distribution of hidden code https://towardsdatascience.com/a-wizards-guide-to-adversarial-autoencoders-part-2- exploring-latent-space-with-adversarial-2d53a6f8a4f9 f9 Adversarial AutoEncoders ▪ Give more details after GAN https://towardsdatascience.com/a-wizards-guide-to-adversarial-autoencoders-part-2- exploring-latent-space-with-adversarial-2d53a6f8a4f9 Another

and introduce techniques like Synthetic Minority Oversampling Technique16 (SMOTE) and Generative Adversarial Network17 (GAN) which can generate synthetic data for images. While SMOTE leverages statistical synthesized. The rightmost is the synthesized image for class 0. 17 Goodfellow, Ian, et al. "Generative adversarial nets." Advances in neural information processing systems 27 (2014). 16 Chawla, Nitesh V., et al over time to be increasingly sophisticated agents. Figure 3-15: Architecture of a Generative Adversarial Network (GAN). It has three phases: discriminator training, generator training and the synthetic

Scientific Reading [Graham Cormode: How NOT to review a paper: the tools and techniques of the adversarial reviewer. SIGMOD Rec. 37(4) 2008] This paper leaves many questions unanswered. Some claims are datasets  Conclusions  Disagree w/ every claim; future work can be dismissed Scientific Reading Adversarial Paper Summary This paper attempts to address the well- studied problem of Graticule Optimization

cloud-hypervisor/cloud-hypervisor 172 1035 1062 915 10 Trusted-AI/adversarial-robustness-toolbox 495.1682899579140 1046 1283 10 Trusted-AI/adversarial- robustness-toolbox 228 305 687 272 11 JanusGraph/janusgraph cloud-hypervisor/cloud-hypervisor 172 1035 1062 915 10 Trusted-AI/adversarial-robustness-toolbox 495.1682899579140 1046 1283 10 Trusted-AI/adversarial- robustness-toolbox 228 305 687 272 11 JanusGraph/janusgraph

Metz 和 S. Chintala, Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks, 2015. 预览版202112 第11章 循环神经网络 人工智能的强力崛起，可能是人类历史上最好的事 情，也可能是最糟糕的事情。−史蒂芬•霍金 实现非常简单，通过在网络层 中插入 Dropout 层即可实现网络连接的随机断开。 12.3.3 Adversarial Auto-Encoder 为了能够方便地从某个已知的先验分布中?(?)采样隐藏变量?，方便利用?(?)来重建输 入，对抗自编码器(Adversarial Auto-Encoder)利用额外的判别器网络(Discriminator，简称 D 网络)来判定降维的隐藏变量 预览版202112 第13章 生成对抗网络 我不能创造的事物，我就还没有完全理解它。−理查 德·費曼 在生成对抗网络(Generative Adversarial Network，简称 GAN)发明之前，变分自编码器 被认为是理论完备，实现简单，使用神经网络训练起来很稳定，生成的图片逼近度也较 高，但是人眼还是可以很轻易地分辨出真实图片与机器生成的图片。

Information Huge State and Action Space Long-Term Planning Temporal and Spatial Reasoning Adversarial Real-time Strategy Multiagent Cooperation StarCraft AI Research and Competitions Classic

▪ Painter or Generator: ▪ Critic or Discriminator https://towardsdatascience.com/generative-adversarial-networks-explained- 34472718707a How to train? https://lilianweng.github.io/lil-log/2017/08/20/from-GAN-to-WGAN

Development began to see operational issues, and usability problems  The relationship is less adversarial and more supportive.  Individuals are cross-trained on each other’s concerns – empathy  Combined

ca/~vincentp/ift3395/lectures/backprop_old.pdf GAN:2014 https://papers.nips.cc/paper/5423-generative-adversarial-nets.pdf BigGAN https://arxiv.org/pdf/1809.11096.pdf Ian Goodfellow ▪ How I fail https://veronikach