Hossein Sarrafzadeh

An incremental learning approach to continuous image change detection

This paper proposes a novel incremental learning based image change detection method capable of detecting changes over image series. Given two images for change detection, an intelligent agent is trained by incremental learning on the source image. As detecting changes to target image, the agent conducts “one-step more” incremental learning on the target image to find its difference against what has been just learned from the source image. For detecting continuously changes to the third image, the agent upgrade its knowledge on the second image by performing incremental learning on top of its current knowledge. For performance evaluation, we performed extensive change detection experiments on both static images and image series. The results show that the proposed approach not only provides consistently accurate image detection, but also demonstrates substantial memory efficiency improvements when compared to existing methods.

UnitecDEAMP: Flow Feature Profiling for Malicious Events Identification in Darknet Space

This paper proposes a traffic decomposition approach called UnitecDEAMP based on flow feature profiling to distinct groups of significant malicious events from background noise in massive historical darknet traffic. Specifically, we segment and extract traffic flows from captured darknet data, categorize the flows according to sets of criteria derived from our traffic behavior assessments. Those criteria will be validated through the followed correlation analysis to guarantee that any redundant criteria be eliminated. Significant events are appraised by combined criteria filtering, including significance regarding volume, significance in terms of time series occurrence and significance regarding variation. To demonstrate the effectiveness of our UnitecDEAMP, real world darknet traffic data sets with twelve months are used for conducting our empirical study. The experimental results show that UnitecDEAMP can effectively select the most significant malicious events.

Ensemble Statistical and Heuristic Models for Unsupervised Word Alignment

Statistical word alignment models need large amounts of training data while they are weak in small-sized corpora. This paper proposes a new approach of an unsupervised hybrid word alignment technique using an ensemble learning method. This algorithm uses three base alignment models in several rounds to generate alignments. The ensemble algorithm uses a weighed scheme for resampling training data and a voting score to consider aggregated alignments. The underlying alignment algorithms used in this study include IBM Model 1, 2 and a heuristic method based on Dice measurement. Our experimental results show that by this approach, the alignment error rate could be improved by at least 15% for the base alignment models.