Deep Learning Security Workshop

Introduction Abstract

Networks and graphs are prevalent in many real world applications such as online social networks, transactions in payment platforms, user-item interactions in recommendation systems and relational information in knowledge bases. The availability of large amount of graph data has posed great new challenges. How to represent such data to capture similarities or differences between involved entities? How to learn predictive models and perform reasoning based on a large amount of such data? Previous deep learning models, such as CNN and RNN, are designed for images and sequences, but they are not applicable to graph data.

In this talk, I will present an embedding framework, called Structure2Vec, for learning representation for graph data in an end-to-end fashion. Structure2Vec provides a unified framework for integrating information from node characteristics, edge features, heterogeneous network structures and network dynamics, and linking them to downstream supervised and unsupervised learning, and reinforcement learning. I will also discuss several applications in security analytics where Structure2Vec leads to significant improvement over previous state-of-the-arts.

Most enterprises provide Web services open to the public and thus are prone to Web attacks. Among these attacks, SQL Injection (SQLi) is a very serious threat, which could lead to catastrophic data leaking and loss. Lots of research efforts have been devoted into SQLi detection areas thereafter and heuristic rule based Web Application Firewall (WAF) was invented and deployed for protections. However, subtle and carefully crafted SQLi attacks can still easily hide themselves from being detected. Deep learning methods were introduced in recent years and they have shown a great strength to learn and describe complex semantic contents. In this talk, we present a CNN based SQLi detection implementation, which also has an ability to propose locations of suspect attack payloads within an URL request. Experimental results show that our implementation outperforms a leading method called Libinjection and our current version Alibaba Cloud WAF.

The proliferation of interconnected sensors all around us opens up new opportunities and challenges in machine learning, security, and authentication. This talk covers our real world experience in developing an implicit authentication platform using deep learning on sensor data and user behavior. Humans are creatures of habit and there are unique behavioral aspects in the way people walk, type, and move that are both unique and consistent. We have developed a solution that can achieve a high level of accuracy (>99.999%) based purely on passive factors, without requiring any conscious user action.

Developing a implicit authentication solution that works in the real world requires a creative application of machine learning as well as new innovations and techniques. We will cover some of the challenges that we have encountered applying deep learning techniques to user authentication in the real world.

Info about poster presentations

I will talk about what machine learning privacy is, and will discuss how and why machine learning models leak information about the individual data records on which they were trained. My quantitative analysis will be based on the fundamental membership inference attacks: given a data record and (black-box) access to a model, determine if a record was in the model’s training set. I will demonstrate how to build such inference attacks on different classification models e.g., trained by commercial “machine learning as a service” providers such as Google and Amazon.

Deep Learning models are vulnerable to adversarial attacks, which can reliably cause the models to misbehave, for example convincing an image classifier that an image of a cat is actually a dog. We will discuss some of the recent attacks and attempts at defending against such attacks, and look at how adversarial attacks may be harnessed to improve the robustness of Deep Learning models.

As a kind of biometrics, fingerprint recognition is so popular and been used all over the world,for example, the mobile phone,the access control system,and so on.As all people think, fingerprint recognition is more secure than others, such as password. And it’s so difficult to clone a fingerprint, for example,use it to unlock the screen of mobile phone. But we have found a way to easily capture and clone a person’s fingerprint.It’s so fast and has a low cost. we well show you how to clone a fingerprint with a paper and a foam pad in a few minutes. In consideration of the subject of this meeting, we will show you the research of how to find a fingerprint use fast RCNN, Fast Region-based Convolutional Networks for object detection. we just want to talk about how to solve security problems with deep learning, and operate in actual situation.

Info about poster presentations