Search

Given a machine learning model and a record, membership attacks determine whether this record was used as part of the model's training dataset. Membership inference can present a risk to private datasets if these datasets are used to train machine learning models and access to the resulting models is open to the public. To construct attack models, multiple shadow models are created that imitate the behaviour of the target model, but for which we know the training datasets and thus the ground truth about membership in these datasets. Attack models are then trained on the labeled inputs and outputs of the shadow models. There is a desideratum to conduct more analysis about this attack and accordingly to provide robust mitigation techniques that will not affect the target model's utility. In this paper, we empirically analyzed this attack from different perspectives related to the number of models and the type of training algorithms. We also proposed and evaluated different mitigation techniques against this type of attack considering different training algorithms of the target model. Our experiments show that the defence strategies mitigate the membership inference attack considerably while preserving the utility of the target model. Finally, we summarized and compared the existing mitigation techniques with our results.

Erroneous data can cause a system to be failed. Though there are several methods for detection and correction, with the increasing amount of errors, it becomes difficult, for both detection, and correction of these erroneous codes. For solving these issues, this paper represents an effective method for solving multiple errors by using Horizontal-Vertical-SuperQueen (HVSQ) parity bits in code. It works with 121 data bits and 44 parity bits. And this method has a higher correction rate with less code overhead and higher code-rate. For these 121 bits of data, we need only 44 redundant bits which, indicate 36.36% of bit overhead and can solve up to 3 bit of errors. It also shows better accuracy in the increased number of errors in data bits.

Comments are supported by various websites and provide a simple approach to increment user involvement. Users can generally comment on different types of media such as: social networks, blogs, forums and news articles. As discussions increasingly move toward online forums, the issue of insulting and abusive comments is becoming prevalent. In addition, a lots of comments are available due to these social media. Hence, it is not feasible for a human moderator to check each comments one by one and flag them as abusive or not abusive. For this reason, an automated classifier which is quick and efficient is necessary to detect such type of comments. To fulfill above purpose, in this paper a Naive Bayes classifier is designed to detect abusive comments expressed in Bangla. Using a training corpus collected from “Youtube.com”, the Naive Bayes classifier is employed to categorize comments as abusive or not abusive. Finally, the performance is evaluated by using 10-fold cross-validation on unprocessed data.

This paper proposes a new method to detect and correct multi bit errors in memory applications using a combination of a clustering approach, Bit-Per-Byte error detection technique, and Majority Logic Decodable (MLD) codes. The likelihood of soft errors accelerates with system complexity, reduction in operational voltages, exponential growth in transistor per chip, increases in clock frequencies, breakdown of memory reliability and device shrinking. Memories are the sensitive part of a computer system. Soft errors in memories may cause an instruction to malfunction. Several techniques are already in practice to mitigate the soft errors. Majority logic decodable codes are proved as effective for memory applications because of their ability to correct a massive number of errors. Since memories are used to hold large number of bits that’s the restraint of Majority logic decodable codes method, so we emphasize on the size of data word in this method. The proposed method aims to detect and correct up to seven bit errors with lesser computational time. It works in an efficient manner in case of adjacent errors which is not possible in Majority logic decodable codes (MLD). It is delineated by Experimental reviews that the proposed approach outperforms existing dominant approach with respect to number of erroneous bit detection and correction, and computational time overhead.

The issue of crowd reckoning has as of late gotten noteworthy consideration in the current era. Given an image of a crowded environment, the objective is to evaluate the density of the crowd, by counting the quantity of individuals that it contains. There are some approaches to solve crowd reckoning problem. They intend to reckon crowd for a medium sized place. But crowd detection at Hajj Pilgrimage site has never been tried before as far we have studied. Hajj Pilgrimage sites are different than most of the places. The aim of this work is to prevent any future stampede at Hajj Pilgrimage sites. This paper proposes a method where head detection technique is used for the crowd reckoning. It uses the segmentation of head regions from still images. The crowd reckoning will determine the possibility of a stampede and can alarm the people to take necessary steps. Assessment of the proposed model demonstrates worthy performance in realistic operating conditions.

In this paper, we present an applied research on designing and developing a recommender system for graduate admission seekers which can help them to choose graduate school matching their entire academic profile. Here we have developed a technique to transform relational database for students' all types of relevant information into a universal database format using academic data of successful students who have already got opportunity to study abroad. After that we have developed an algorithm for grad school recommender system which can calculate similarity between training and test data set based on weighted scores using mean squared deviation similarity metrics. We have used K-nearest Neighbor algorithm for calculating top N similar users for the test users and recommend Top K universities to users from N similar users. Finally our proposed recommender system will recommend list of universities to apply for graduate admission to pursue higher study abroad with funding.

Soft error tolerance is a matter of concern for system reliability nowadays. The likelihood of soft errors increase with system complexity, reduction in operational voltages, exponential growth in transistors per chip, increases in clock frequencies and device shrinking. As the memory bit-cell area is condensed, single event upset that would have formerly despoiled only a single bit-cell are now proficient of upsetting multiple contiguous memory bit-cells per particle strike. While these error types are beyond the error handling capabilities of the frequently used error correction codes (ECCs) for single bit, the overhead associated with moving to more sophisticated codes for multi-bit errors is considered to be too costly. To address this issue, this paper presents a new approach to detect and correct multi-bit soft error by using Horizontal-Vertical-Diagonal Blocks (HVDB) parity bits with a higher reliability.