Biblio

Data poisoning is a type of adversarial attack on training data where an attacker manipulates a fraction of data to degrade the performance of machine learning model. There are several known defensive mechanisms for handling offline attacks, however defensive measures for online learning, where data points arrive sequentially, have not garnered similar interest. In this work, we propose a defense mechanism to minimize the degradation caused by the poisoned training data on a learner's model in an online setup. Our proposed method utilizes an influence function which is a classic technique in robust statistics. Further, we supplement it with the existing data sanitization methods for filtering out some of the poisoned data points. We study the effectiveness of our defense mechanism on multiple datasets and across multiple attack strategies against an online learner.

The graphical pattern unlock scheme which requires users to connect a minimum of 4 nodes on 3X3 grid is one of the most popular authentication mechanism on mobile devices. However prior research suggests that users' pattern choices are highly biased and hence vulnerable to guessing attacks. Moreover, 3X3 pattern choices are devoid of features such as longer stroke lengths, direction changes and intersections that are considered to be important in preventing shoulder-surfing attacks. We attribute these insecure practices to the geometry of the grid and its complicated drawing rules which prevent users from realising the full potential of graphical passwords. In this paper, we propose and explore an alternate circular layout referred to as Pass-O which unlike grid layout allows connection between any two nodes, thus simplifying the pattern drawing rules. Consequently, Pass-O produces a theoretical search space of 9,85,824, almost 2.5 times greater than 3X3 grid layout. We compare the security of 3X3 and Pass-O patterns theoretically as well as empirically. Theoretically, Pass-O patterns are uniform and have greater visual complexity due to large number of intersections. To perform empirical analysis, we conduct a large-scale web-based user study and collect more than 1,23,000 patterns from 21,053 users. After examining user-chosen 3X3 and Pass-O patterns across different metrics such as pattern length, stroke length, start point, end point, repetitions, number of direction changes and intersections, we find that Pass-O patterns are much more secure than 3X3 patterns.