Trojan signatures, as described by Fields et al. (2021), are noticeable differences in the distribution of the trojaned class parameters (weights) and the non-trojaned class parameters of the trojaned model, that can be used to detect the trojaned model. Fields et al. (2021) found trojan signatures in computer vision classification tasks with image models, such as, Resnet, WideResnet, Densenet, and VGG. In this paper, we investigate such signatures in the classifier layer parameters of large language models of source code. Our results suggest that trojan signatures could not generalize to LLMs of code. We found that trojaned code models are stubborn, even when the models were poisoned under more explicit settings (finetuned with pre-trained weights frozen). We analyzed nine trojaned models for two binary classification tasks: clone and defect detection. To the best of our knowledge, this is the first work to examine weight-based trojan signature revelation techniques for large-language models of code and furthermore to demonstrate that detecting trojans only from the weights in such models is a hard problem.
Machine learning models are widely used but can also often be wrong. Users would benefit from a reliable indication of whether a given output from a given model should be trusted, so a rational decision can be made whether to use the output or not. For example, outputs can be associated with a confidence measure; if this confidence measure is strongly associated with likelihood of correctness, then the model is said to be well-calibrated. In this case, for example, high-confidence outputs could be safely accepted, and low-confidence outputs rejected. Calibration has so far been studied in non-generative (e.g., classification) settings, especially in Software Engineering. However, generated code can quite often be wrong: Developers need to know when they should e.g., directly use, use after careful review, or discard model-generated code; thus Calibration is vital in generative settings. However, the notion of correctness of generated code is non-trivial, and thus so is Calibration. In this paper we make several contributions. We develop a framework for evaluating the Calibration of code-generating models. We consider several tasks, correctness criteria, datasets, and approaches, and find that by and large generative code models are not well-calibrated out of the box. We then show how Calibration can be improved, using standard methods such as Platt scaling. Our contributions will lead to better-calibrated decision-making in the current use of code generated by language models, and offers a framework for future research to further improve calibration methods for generative models in Software Engineering.
Although deep neural models substantially reduce the overhead of feature engineering, the features readily available in the inputs might significantly impact training cost and the performance of the models. In this paper, we explore the impact of an unsuperivsed feature enrichment approach based on variable roles on the performance of neural models of code. The notion of variable roles (as introduced in the works of Sajaniemi et al. [Refs. 1,2]) has been found to help students' abilities in programming. In this paper, we investigate if this notion would improve the performance of neural models of code. To the best of our knowledge, this is the first work to investigate how Sajaniemi et al.'s concept of variable roles can affect neural models of code. In particular, we enrich a source code dataset by adding the role of individual variables in the dataset programs, and thereby conduct a study on the impact of variable role enrichment in training the Code2Seq model. In addition, we shed light on some challenges and opportunities in feature enrichment for neural code intelligence models.
Finding important features that contribute to the prediction of neural models is an active area of research in explainable AI. Neural models are opaque and finding such features sheds light on a better understanding of their predictions. In contrast, in this work, we present an inverse perspective of distractor features: features that cast doubt about the prediction by affecting the model's confidence in its prediction. Understanding distractors provide a complementary view of the features' relevance in the predictions of neural models. In this paper, we apply a reduction-based technique to find distractors and provide our preliminary results of their impacts and types. Our experiments across various tasks, models, and datasets of code reveal that the removal of tokens can have a significant impact on the confidence of models in their predictions and the categories of tokens can also play a vital role in the model's confidence. Our study aims to enhance the transparency of models by emphasizing those tokens that significantly influence the confidence of the models.
Neural code intelligence (CI) models are opaque black-boxes and offer little insight on the features they use in making predictions. This opacity may lead to distrust in their prediction and hamper their wider adoption in safety-critical applications. Recently, input program reduction techniques have been proposed to identify key features in the input programs to improve the transparency of CI models. However, this approach is syntax-unaware and does not consider the grammar of the programming language. In this paper, we apply a syntax-guided program reduction technique that considers the grammar of the input programs during reduction. Our experiments on multiple models across different types of input programs show that the syntax-guided program reduction technique is faster and provides smaller sets of key tokens in reduced programs. We also show that the key tokens could be used in generating adversarial examples for up to 65% of the input programs.
The code intelligence (CI) models are often black-box and do not offer any insights on the input features that they learn for making correct predictions. This opacity may lead to distrust in their prediction and hamper their wider adoption in safety-critical applications. In recent, the program reduction technique is widely being used to identify key input features in order to explain the prediction of CI models. The approach removes irrelevant parts from an input program and keeps the minimal snippets that a CI model needs to maintain its prediction. However, the state-of-the-art approaches mainly use a syntax-unaware program reduction technique that does not follow the syntax of programs, which adds significant overhead to the reduction of input programs and explainability of models. In this paper, we apply a syntax-guided program reduction technique that follows the syntax of input programs during reduction. Our experiments on multiple models across different types of input programs show that the syntax-guided program reduction technique significantly outperforms the syntax-unaware program reduction technique in reducing the size of input programs. Extracting key input features from reduced programs reveals that the syntax-guided reduced programs contain more label-specific key input features and are more vulnerable to adversarial transformation when renaming the key tokens in programs. These label-specific key input features may help to understand the reasoning of models' prediction from different perspectives and increase the trustworthiness to correct classification given by CI models.
There are several approaches to encode source code in the input vectors of neural models. These approaches attempt to include various syntactic and semantic features of input programs in their encoding. In this paper, we investigate Code2Snapshot, a novel representation of the source code that is based on the snapshots of input programs. We evaluate several variations of this representation and compare its performance with state-of-the-art representations that utilize the rich syntactic and semantic features of input programs. Our preliminary study on the utility of Code2Snapshot in the code summarization task suggests that simple snapshots of input programs have comparable performance to the state-of-the-art representations. Interestingly, obscuring the input programs have insignificant impacts on the Code2Snapshot performance, suggesting that, for some tasks, neural models may provide high performance by relying merely on the structure of input programs.
Deep Neural Networks (DNN) are increasingly commonly used in software engineering and code intelligence tasks. These are powerful tools that are capable of learning highly generalizable patterns from large datasets through millions of parameters. At the same time, training DNNs means walking a knife's edges, because their large capacity also renders them prone to memorizing data points. While traditionally thought of as an aspect of over-training, recent work suggests that the memorization risk manifests especially strongly when the training datasets are noisy and memorization is the only recourse. Unfortunately, most code intelligence tasks rely on rather noise-prone and repetitive data sources, such as GitHub, which, due to their sheer size, cannot be manually inspected and evaluated. We evaluate the memorization and generalization tendencies in neural code intelligence models through a case study across several benchmarks and model families by leveraging established approaches from other fields that use DNNs, such as introducing targeted noise into the training dataset. In addition to reinforcing prior general findings about the extent of memorization in DNNs, our results shed light on the impact of noisy dataset in training.