Microarchitectural code analyzers, i.e., tools that estimate the throughput of machine code basic blocks, are important utensils in the tool belt of performance engineers. Recent tools like llvm-mca, uiCA, and Ithemal use a variety of techniques and different models for their throughput predictions. When put to the test, it is common to see these state-of-the-art tools give very different results. These inconsistencies are either errors, or they point to different and rarely documented assumptions made by the tool designers. In this paper, we present AnICA, a tool taking inspiration from differential testing and abstract interpretation to systematically analyze inconsistencies among these code analyzers. Our evaluation shows that AnICA can summarize thousands of inconsistencies in a few dozen descriptions that directly lead to high-level insights into the different behavior of the tools. In several case studies, we further demonstrate how AnICA automatically finds and characterizes known and unknown bugs in llvm-mca, as well as a quirk in AMD's Zen microarchitectures.
相關內容
Ransomware is a malicious class of software that utilises encryption to implement an attack on system availability. The target's data remains encrypted and is held captive by the attacker until a ransom demand is met. A common approach used by many crypto-ransomware detection techniques is to monitor file system activity and attempt to identify encrypted files being written to disk, often using a file's entropy as an indicator of encryption. However, often in the description of these techniques, little or no discussion is made as to why a particular entropy calculation technique is selected or any justification given as to why one technique is selected over the alternatives. The Shannon method of entropy calculation is the most commonly-used technique when it comes to file encryption identification in crypto-ransomware detection techniques. Overall, correctly encrypted data should be indistinguishable from random data, so apart from the standard mathematical entropy calculations such as Chi-Square, Shannon Entropy and Serial Correlation, the test suites used to validate the output from pseudo-random number generators would also be suited to perform this analysis. he hypothesis being that there is a fundamental difference between different entropy methods and that the best methods may be used to better detect ransomware encrypted files. The paper compares the accuracy of 53 distinct tests in being able to differentiate between encrypted data and other file types. The testing is broken down into two phases, the first phase is used to identify potential candidate tests, and a second phase where these candidates are thoroughly evaluated. To ensure that the tests were sufficiently robust, the NapierOne dataset is used. This dataset contains thousands of examples of the most commonly used file types, as well as examples of files that have been encrypted by crypto-ransomware.
Deep neural networks have shown excellent prospects in speech separation tasks. However, obtaining good results while keeping a low model complexity remains challenging in real-world applications. In this paper, we provide a bio-inspired efficient encoder-decoder architecture by mimicking the brain's top-down attention, called TDANet, with decreased model complexity without sacrificing performance. The top-down attention in TDANet is extracted by the global attention (GA) module and the cascaded local attention (LA) layers. The GA module takes multi-scale acoustic features as input to extract global attention signal, which then modulates features of different scales by direct top-down connections. The LA layers use features of adjacent layers as input to extract the local attention signal, which is used to modulate the lateral input in a top-down manner. On three benchmark datasets, TDANet consistently achieved competitive separation performance to previous state-of-the-art (SOTA) methods with higher efficiency. Specifically, TDANet's multiply-accumulate operations (MACs) are only 5\% of Sepformer, one of the previous SOTA models, and CPU inference time is only 10\% of Sepformer. In addition, a large-size version of TDANet obtained SOTA results on three datasets, with MACs still only 10\% of Sepformer and the CPU inference time only 24\% of Sepformer. Our study suggests that top-down attention can be a more efficient strategy for speech separation.
Machine learning (ML) algorithms are gaining increased importance in many academic and industrial applications, and such algorithms are, accordingly, becoming common components in computer science curricula. Learning ML is challenging not only due to its complex mathematical and algorithmic aspects, but also due to a) the complexity of using correctly these algorithms in the context of real-life situations and b) the understanding of related social and ethical issues. Cognitive biases are phenomena of the human brain that may cause erroneous perceptions and irrational decision-making processes. As such, they have been researched thoroughly in the context of cognitive psychology and decision making; they do, however, have important implications for computer science education as well. One well-known cognitive bias, first described by Kahneman and Tversky, is the base rate neglect bias, according to which humans fail to consider the base rate of the underlying phenomena when evaluating conditional probabilities. In this paper, we explore the expression of the base rate neglect bias in ML education. Specifically, we show that about one third of students in an Introduction to ML course, from varied backgrounds (computer science students and teachers, data science, engineering, social science and digital humanities), fail to correctly evaluate ML algorithm performance due to the base rate neglect bias. This failure rate should alert educators and promote the development of new pedagogical methods for teaching ML algorithm performance.
Side-Channel Attacks (SCAs) exploit data correla-tion in signals leaked from devices to jeopardize confidentiality. Locating and synchronizing segments of interest in traces from Cryptographic Processes (CPs) is a key step of the attack. The most common method consists in generating a trigger signal to indicate to the attacker the start of a CP. This paper proposes a method called Virtual Triggering (VT) that removes the need for the trigger signal and automates trace segmentation. When the time between repetitions is not constant, further trace alignment techniques are required. Building on VT, we propose a simple method to learn representative segment templates from a profiling device similar to the victim, and to automatically locate and pull out these segments from other victim devices using simple pattern recognition. We evaluate VT on screaming channel attacks [1], which initially used a Frequency Component (FC) known to appear at a single time in leaked signals, as a trigger to segment traces. We demonstrate that VT not only performs equivalently to FC on a standard attack scenario, but we also show how using VT with the automatic pullout technique improves the attack efficiency and enables more realistic attack scenarios. Thanks to VT, screaming channel attacks can now: (1) succeed with only half of the segments collected compared to the FC trigger from the original attack; and (2) absorb time variations between CPs.
Supplier selection and order allocation (SSOA) are key strategic decisions in supply chain management which greatly impact the performance of the supply chain. The SSOA problem has been studied extensively but the lack of attention paid to scalability presents a significant gap preventing adoption of SSOA algorithms by industrial practitioners. This paper presents a novel real-time large-scale industrial SSOA problem, which involves a multi-item, multi-supplier environment with dual-sourcing and penalty constraints across two-tiers of a supply chain of a manufacturing company. The problem supports supplier preferences to work with other suppliers through bidding. This is the largest scale studied so far in literature, and needs to be solved in a real-time auction environment, making computational complexity a key issue. Furthermore, order allocation needs to be undertaken on both supply tiers, with dynamically presented constraints where non-preferred allocation may results in penalties by the suppliers. We subsequently propose Mixed Integer Programming models for individual-tiers as well as an integrated problem, which are complex due to NP-hard nature. The use case allows us to highlight how problem formulation, modelling and choice of modelling can help reduce complexity using Mathematical Programming (MP) and Genetic Algorithm (GA) approaches. The results show an interesting observation that MP outperforms GA to solve the individual-tiers problem as well as the integrated problem. Sensitivity analysis is presented for sourcing strategy, penalty threshold and penalty factor. The developed model was successfully deployed in a supplier conference which helped in significant procurement cost reductions to the manufacturing company.
Discovering new intents is of great significance to establishing Bootstrapped Task-Oriented Dialogue System. Most existing methods either lack the ability to transfer prior knowledge in the known intent data or fall into the dilemma of forgetting prior knowledge in the follow-up. More importantly, these methods do not deeply explore the intrinsic structure of unlabeled data, so they can not seek out the characteristics that make an intent in general. In this paper, starting from the intuition that discovering intents could be beneficial to the identification of the known intents, we propose a probabilistic framework for discovering intents where intent assignments are treated as latent variables. We adopt Expectation Maximization framework for optimization. Specifically, In E-step, we conduct discovering intents and explore the intrinsic structure of unlabeled data by the posterior of intent assignments. In M-step, we alleviate the forgetting of prior knowledge transferred from known intents by optimizing the discrimination of labeled data. Extensive experiments conducted in three challenging real-world datasets demonstrate our method can achieve substantial improvements.
Users today expect more security from services that handle their data. In addition to traditional data privacy and integrity requirements, they expect transparency, i.e., that the service's processing of the data is verifiable by users and trusted auditors. Our goal is to build a multi-user system that provides data privacy, integrity, and transparency for a large number of operations, while achieving practical performance. To this end, we first identify the limitations of existing approaches that use authenticated data structures. We find that they fall into two categories: 1) those that hide each user's data from other users, but have a limited range of verifiable operations (e.g., CONIKS, Merkle2, and Proofs of Liabilities), and 2) those that support a wide range of verifiable operations, but make all data publicly visible (e.g., IntegriDB and FalconDB). We then present TAP to address the above limitations. The key component of TAP is a novel tree data structure that supports efficient result verification, and relies on independent audits that use zero-knowledge range proofs to show that the tree is constructed correctly without revealing user data. TAP supports a broad range of verifiable operations, including quantiles and sample standard deviations. We conduct a comprehensive evaluation of TAP, and compare it against two state-of-the-art baselines, namely IntegriDB and Merkle2, showing that the system is practical at scale.
Today, an increasing number of Adaptive Deep Neural Networks (AdNNs) are being used on resource-constrained embedded devices. We observe that, similar to traditional software, redundant computation exists in AdNNs, resulting in considerable performance degradation. The performance degradation is dependent on the input and is referred to as input-dependent performance bottlenecks (IDPBs). To ensure an AdNN satisfies the performance requirements of resource-constrained applications, it is essential to conduct performance testing to detect IDPBs in the AdNN. Existing neural network testing methods are primarily concerned with correctness testing, which does not involve performance testing. To fill this gap, we propose DeepPerform, a scalable approach to generate test samples to detect the IDPBs in AdNNs. We first demonstrate how the problem of generating performance test samples detecting IDPBs can be formulated as an optimization problem. Following that, we demonstrate how DeepPerform efficiently handles the optimization problem by learning and estimating the distribution of AdNNs' computational consumption. We evaluate DeepPerform on three widely used datasets against five popular AdNN models. The results show that DeepPerform generates test samples that cause more severe performance degradation (FLOPs: increase up to 552\%). Furthermore, DeepPerform is substantially more efficient than the baseline methods in generating test inputs(runtime overhead: only 6-10 milliseconds).
Since hardware resources are limited, the objective of training deep learning models is typically to maximize accuracy subject to the time and memory constraints of training and inference. We study the impact of model size in this setting, focusing on Transformer models for NLP tasks that are limited by compute: self-supervised pretraining and high-resource machine translation. We first show that even though smaller Transformer models execute faster per iteration, wider and deeper models converge in significantly fewer steps. Moreover, this acceleration in convergence typically outpaces the additional computational overhead of using larger models. Therefore, the most compute-efficient training strategy is to counterintuitively train extremely large models but stop after a small number of iterations. This leads to an apparent trade-off between the training efficiency of large Transformer models and the inference efficiency of small Transformer models. However, we show that large models are more robust to compression techniques such as quantization and pruning than small models. Consequently, one can get the best of both worlds: heavily compressed, large models achieve higher accuracy than lightly compressed, small models.
To address the sparsity and cold start problem of collaborative filtering, researchers usually make use of side information, such as social networks or item attributes, to improve recommendation performance. This paper considers the knowledge graph as the source of side information. To address the limitations of existing embedding-based and path-based methods for knowledge-graph-aware recommendation, we propose Ripple Network, an end-to-end framework that naturally incorporates the knowledge graph into recommender systems. Similar to actual ripples propagating on the surface of water, Ripple Network stimulates the propagation of user preferences over the set of knowledge entities by automatically and iteratively extending a user's potential interests along links in the knowledge graph. The multiple "ripples" activated by a user's historically clicked items are thus superposed to form the preference distribution of the user with respect to a candidate item, which could be used for predicting the final clicking probability. Through extensive experiments on real-world datasets, we demonstrate that Ripple Network achieves substantial gains in a variety of scenarios, including movie, book and news recommendation, over several state-of-the-art baselines.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191