Search

The Transformers architecture has recently emerged as a revolutionary paradigm in the field of deep learning, particularly excelling in Natural Language Processing (NLP) and Computer Vision (CV) applications. Despite its success, the security implications of Transformers have not been comprehensively explored, encompassing a broad spectrum of both hardware and software vulnerabilities. This paper aims to address this critical gap by conducting an extensive exploration of security challenges confronting Transformers from both software and hardware perspectives. While software-related concerns like adversarial attacks, private inference, and watermarking have been studied, the paper sheds light on previously underexplored hardware vulnerabilities such as trojans and side-channel attacks. By unraveling the intricacies of these hardware threats, the study aims to contribute to a comprehensive understanding of Transformer security. It presents an in-depth analysis of recent advancements in the security of Transformers. Additionally, it outlines existing challenges and forecasts future research trends, offering insights for researchers and practitioners aiming for the secure and resilient design and deployment of Transformers. The survey categorizes different attacks and defenses related to Transformers, helping researchers identify gaps and opportunities in this area. Furthermore, it defines a roadmap for a unified security framework, serving as a foundational starting point for developers seeking to implement robust security measures.

Evapotranspiration, as a combination of evaporation and transpiration of water vapour, is a primary component of global hydrological cycles. It accounts for significant loss of soil moisture from the earth to the atmosphere. Reliable methods to monitor and forecast evapotranspiration are required for decision-making. Reference evapotranspiration, denoted as ET, is a major parameter that is useful in quantifying soil moisture in a cropping system. This article aims to design a multi-stage deep learning hybrid Long Short-Term Memory (LSTM) predictive model that is coupled with Multivariate Empirical Mode Decomposition (MEMD) and Boruta-Random Forest (Boruta) algorithms to forecast ET in the drought-prone regions (i.e., Gatton, Fordsdale, Cairns) of Queensland, Australia. Daily data extracted from NASA’s Goddard Online Interactive Visualization and Analysis Infrastructure (GIOVANNI) and Scientific Information for Land Owners (SILO) repositories over 2003–2011 are used to build the proposed multi-stage deep learning hybrid model, i.e., MEMD-Boruta-LSTM, and the model’s performance is compared against competitive benchmark models such as hybrid MEMD-Boruta-DNN, MEMD-Boruta-DT, and a standalone LSTM, DNN and DT model. The test MEMD-Boruta-LSTM hybrid model attained the lowest Relative Root Mean Square Error (≤17%), Absolute Percentage Bias (≤12.5%)and the highest Kling-Gupta Efficiency (≥0.89%) relative to benchmark models for all study sites. The proposed multi-stage deep hybrid MEMD-Boruta-LSTM model also outperformed all other benchmark models in terms of predictive efficacy, demonstrating its usefulness in the forecasting of the daily ET dataset. This MEMD-Boruta-LSTM hybrid model could therefore be employed in practical environments such as irrigation management systems to estimate evapotranspiration or to forecast ET.