Sentiment analysis (SA) is crucial for many NLP applications across various domains. While Arabic is one of the world’s major languages, high-quality NLP models developed for standard Arabic often underperform on regional dialects like the Saudi Dialect (SD) due to a lack of SD-specific training data. This paper presents a novel approach to adapting a high-resource language model, AraBERT, for low-resource dialect sentiment analysis by combining minimal SD data collection with generative AI. In the absence of openly accessible SD datasets, we augmented a small amount of collected SD data with GPT-generated SD data to fine-tune AraBERT for sentiment analysis in SD. Our contributions include (1) demonstrating the feasibility of low-effort data collection of a low-resource dialect for adapting existing high-resource NLP models and (2) leveraging GPT-generated data to augment collected data to enhance a high-resource language model for sentiment classification in a low-resource dialect, achieving significant improvements over the pre-trained high-resource model. These two contributions imply a potentially replicable approach that can serve as a template for future research in other low-resource NLP tasks. This paper presents a promising solution for enhancing model performance in low-resource dialects and has implications for similar under-resourced languages.

Double Arbiter PUFs (DAPUFs) were developed as a variant to XOR PUFs to improve resilience against machine learning attacks. A recent study on DAPUFs of sizes up to 4-1 DAPUFs showed that all examined DAPUFs were vulnerable to machine learning attacks when attackers have access to a large number of challenge-response pairs (CRPs) [1], [10]. In this paper, we implemented the 5-1 DAPUF on field programmable gate arrays (FPGAs), larger than all previously implemented DAPUFs, and carried out performance evaluations of 5-1 DAPUFs on various properties including response randomness, uniqueness, stability, and security vulnerability. Experimental study on 5-1 DAPUFs shows that responses from the same 5-1 DAPUF circuit to different challenges are adequately highly distinguishable from each other while responses generated on different devices to the same challenges are different enough. 5-1 DAPUF also records the highest randomness among all tested sizes of DAPUFs. However, the stability issue is exacerbated in 5-1 DAPUF, a drawback that is also revealed in earlier studies of DAPUFs. Machine learning attack experiments show that 5-1 DAPUF is more resilient than other DAPUFs, but its responses could still be modeled when an attacker is able to accumulate a large number of CRPs.

Security is important for the functioning of Internet-of-Things (IoTs). Many IoT devices are closely physically approachable by the crowd, making cryptographic key-based security protocols vulnerable to side-channel attacks. Physical Unclonable Functions (PUF) are emerging as a promising keyless solution by utilizing inherent variations of integrated circuits (ICs) to produce different responses from different devices. While physically unclonable, some PUFs were reported to be mathematically clonable by machine learning (ML) methods. XOR PUFs are a group of PUFs mathematically clonable when a large number of challenge-response pairs (CRPs) are available to attackers. Double Arbiter PUFs (DAPUFs) were developed for increased security against machine learning attacks over XOR PUFs, and studies showed that DAPUFs are highly secure against attacks using Support Vector Machine (SVM). In this paper, we investigate how secure DAPUFs are against neural network-based attack methods and compare the DAPUFs' performance with that of XOR PUFs. The results confirm with existing studies on DAPUFs' higher security when compared with XOR PUFs, but also discovered that DAPUFs are not secure against neural network-based attacks if attackers can obtain a large number of CRPs, revealing a security vulnerability of those DAPUFs we examined in this study.

Physical Unclonable Functions (PUFs), leveraging integrated circuits' manufacturing variations to produce responses unique for individual devices, are emerging as a promising class of security hardware primitives. Implementable with simplistic circuits and requiring low operation energy, PUFs are particularly suitable for resource-constrained systems. An important part of security research is to discover all possible security risks. Such information is useful for PUF developers to design new PUFs to overcome existing risks as well as for PUF-utilizing application developers to avoid vulnerable PUFs. While physically unclonable, some PUFs have been found to be mathematically clonable by machine learning methods which can accurately predict the responses of PUFs. Mathematical clonability allows attackers to develop malicious software to impersonate PUF-embedded devices by producing the same responses PUFs would give. Existing studies on machine learning attack of PUFs have not found vulnerability of large XOR PUFs with component-differential challenges. We believe that the high dimensionality of the challenge space of such PUFs is the underlying reason for the difficulty of machine learning attacks. In this paper, we introduce a PUF-architecture-tailored subspace prelearning-based attack method that can learn the responses of such XOR PUFs fast and accurately, revealing a vulnerability of these XOR PUFs if the PUF has an interface conforming to the way challenge-response data are accessed for the subspace prelearning-based attack method.

The increasing use of big datasets by analytics applications for higher predictive power leads to higher processing overhead, and the overhead becomes more substantial when datasets are larger than memory capacity. In this paper, we focus on reducing I/O overhead for big data machine learning procedures, including both unsupervised and supervised learning. While I/O data are, in general, not reducible in well-developed applications, our approach to I/O overhead reduction is to overlap I/O's with computations so that when an application is performing an I/O, other useful computation is also processed. To this end, we develop an I/O latency-hiding (LaHiIO) strategy and an enabling easy-to-use API, a wrapper of existing asynchronous I/O operations, by hiding away features not likely needed for general data analytics applications and keeping only those necessary for computation-I/O overlapping. By doing so, we aim to increase the use of computation-I/O overlapping in big data applications by a broad range of developers who could be physicists, chemists, biologists, engineers, but not necessarily system programming experts. We apply the LaHiIO strategy to clustering and neural network procedures, the common choices for unsupervised and supervised learning, resulting in significant performance enhancement from about 10% to 150%, indicating the effectiveness of the LaHiIO strategy and its enabling user-friendly API for big data machine learning applications.