Task-oriented dialogue systems often fail before the final breakdown is obvious, but most evaluation only measures failure after the conversation has already gone wrong. We present TRACER, a method for early failure detection in task-oriented dialogue. TRACER predicts from a partial dialogue whether the full conversation will eventually fail by combining simple trajectory signals from belief-state changes with text representations of the evolving dialogue state. We evaluate the method in both oracle and generated belief-state settings, and test how well it works when only 25%, 50%, 75%, or 100% of the dialogue is visible. Across these settings, TRACER detects useful failure signals well before the end of the conversation and outperforms heuristic, classical, and single-stream baselines. These results suggest that early failure detection can provide a practical warning signal for dialogue systems before the interaction fully breaks down. Source code can be found here: https://github.com/erfan-nourbakhsh/TRACER.

As mobile technology has advanced, individuals have started relying on their smartphones to conduct more of their everyday tasks. From playing games or streaming media to social networking and banking, apps on a user's device may have access to the most sensitive information on the device. Privacy policies are designed to inform users of such data practices so that they can make reasonable decisions when using the app. However, an app's true behavior may not always align with the statements in a privacy policy. In this work, we divide our study into two components and compare the viability of various large language models (LLMs): methods for extracting and summarizing privacy policy data practices, or information-type extraction and action-verb extraction; and methods for measuring whether the policy acknowledges the interaction with certain information (sensitive data) compared to identified methods within its app's source code. Fine-tuning GPT-3.5 Turbo delivers a higher average F1-score for both action verb extraction (0.50) and information-type extraction (0.84) compared to other LLMs. ChatGPT outperforms other language models in traditional semantic similarity, providing a consistently high performance, including the highest F1-score (0.52) for this task. Our approaches demonstrate that these LLMs are viable in performing such tasks and additionally that pre-trained instruction-based LLMs are capable of identifying the complex relationships between policies and source code.

Mobile apps are widely used and often process users’ sensitive data. Many taint analysis tools have been applied to analyze sensitive information flows and report data leaks in apps. These tools require a list of sources (where sensitive data is accessed) as input, and researchers have constructed such lists within the Android platform by identifying Android API methods that allow access to sensitive data. However, app developers may also define methods or use third-party library’s methods for accessing data. It is difficult to collect such source methods, because they are unique to the apps, and there are a large number of third-party libraries available on the market that evolve over time. To address this problem, we propose DAISY, a Dynamic-Analysis-Induced Source discoverY approach for identifying methods that return sensitive information from apps and third-party libraries. Trained on an automatically labeled dataset of methods and their calling context, DAISY identifies sensitive methods in unseen apps. We evaluated DAISY on real-world apps, and the results show that DAISY can achieve an overall precision of 77.9% when reporting the most confident results. Most of the identified sources and leaks cannot be detected by existing technologies.

Privacy policies are legal documents containing application data practices. These documents are well-established sources of requirements in software engineering. However, privacy policies are written in natural language, thus subject to ambiguity and abstraction. Eliciting requirements from privacy policies is a challenging task as these ambiguities can result in more than one interpretation of a given information type (e.g., ambiguous information type "device information" in the statement "we collect your device information"). To address this challenge, we propose an automated approach to infer semantic relations among information types and construct an ontology to guide requirements authors in the selection of the most appropriate information type terms. Our solution utilizes word embeddings and Convolutional Neural Networks (CNN) to classify information type pairs as either hypernymy, synonymy, or unknown. We evaluate our model on a manually-built ontology, yielding predictions that identify hypernymy relations in information type pairs with 0.904 F-1 score, suggesting a large reduction in effort required for ontology construction.

Static taint analyses are widely-applied techniques to detect taint flows in software systems. Although they are theoretically conservative and de-signed to detect all possible taint flows, static taint analyses almost always exhibit false negatives due to a variety of implementation limitations. Dynamic programming language features, inaccessible code, and the usage of multiple programming languages in a software project are some of the major causes. To alleviate this problem, we developed a novel approach, DySTA, which uses dynamic taint analysis results as additional sources for static taint analysis. However, naïvely adding sources causes static analysis to lose context sensitivity and thus produce false positives. Thus, we developed a hybrid context matching algorithm and corresponding tool, ConDySTA, to preserve context sensitivity in DySTA. We applied REPRODROID [1], a comprehensive benchmarking framework for Android analysis tools, to evaluate ConDySTA. The results show that across 28 apps (1) ConDySTA was able to detect 12 out of 28 taint flows which were not detected by any of the six state-of-the-art static taint analyses considered in ReproDroid, and (2) ConDySTA reported no false positives, whereas nine were reported by DySTA alone. We further applied ConDySTA and FlowDroid to 100 top Android apps from Google Play, and ConDySTA was able to detect 39 additional taint flows (besides 281 taint flows found by FlowDroid) while preserving the context sensitivity of FlowDroid.

Mobile application (app) developers commonly utilize analytic services to analyze their app users' behavior to support debugging, improve service quality, and facilitate advertising. Anonymization and aggregation can reduce the sensitivity of such behavioral data, therefore analytic services often encourage the use of such protections. However, these protections are not directly enforced so it is possible for developers to misconfigure the analytic services and expose personal information, which may cause greater privacy risks. Since people use apps in many aspects of their daily lives, such misconfigurations may lead to the leaking of sensitive personal information such as a users' real-time location, health data, or dating preferences. To study this issue and identify potential privacy risks due to such misconfigurations, we developed a semi-automated approach, Privacy-Aware Analytics Misconfiguration Detector (PAMDroid), which enables our empirical study on mis-configurations of analytic services. This paper describes a study of 1,000 popular apps using top analytic services in which we found misconfigurations in 120 apps. In 52 of the 120 apps, misconfigurations lead to a violation of either the analytic service providers' terms of service or the app's own privacy policy.

Augmented Reality (AR) is an emerging technique that enriches real environment with virtual information objects. Despite its wide application scenarios, AR techniques also raise concerns on its dependability, especially on the privacy protection of the users and of the people appearing in users' eyesight. In our research, we performed a case study on the mostly popular augmented reality Android app: Google Translate. In this paper, we report our major findings in the case study, and propose potential mechanism to detect unnecessary privacy leaks in Android augmented reality apps.

The Android mobile platform supports billions of devices across more than 190 countries around the world. This popularity coupled with user data collection by Android apps has made privacy protection a well-known challenge in the Android ecosystem. In practice, app producers provide privacy policies disclosing what information is collected and processed by the app. However, it is difficult to trace such claims to the corresponding app code to verify whether the implementation is consistent with the policy. Existing approaches for privacy policy alignment focus on information directly accessed through the Android platform (e.g., location and device ID), but are unable to handle user input, a major source of private information. In this paper, we propose a novel approach that automatically detects privacy leaks of user-entered data for a given Android app and determines whether such leakage may violate the app's privacy policy claims. For evaluation, we applied our approach to 120 popular apps from three privacy-relevant app categories: finance, health, and dating. The results show that our approach was able to detect 21 strong violations and 18 weak violations from the studied apps.

Many access control patterns, both positive and negative, have been identified in the past. However, there is little research describing how to leverage those patterns for the detection of access control bugs in code. Many software bug detection models and frameworks for access control exist, however most of these approaches and tools are process-based and suffer from many limitations. We propose a framework to detect access control bugs based on code pattern detection. Our framework will mine and generate bug patterns, detect those patterns in code, and calculate a vulnerability measure of software. Based on our knowledge we are the first pattern-based model for the detection and measurement of bugs in software. As a proof of concept, we perform a case study of the relational database access control pattern "Improper Authorization''.

Mobile applications frequently access sensitive personal information to meet user or business requirements. Because such information is sensitive in general, regulators increasingly require mobile-app developers to publish privacy policies that describe what information is collected. Furthermore, regulators have fined companies when these policies are inconsistent with the actual data practices of mobile apps. To help mobile-app developers check their privacy policies against their apps' code for consistency, we propose a semi-automated framework that consists of a policy terminology-API method map that links policy phrases to API methods that produce sensitive information, and information flow analysis to detect misalignments. We present an implementation of our framework based on a privacy-policy-phrase ontology and a collection of mappings from API methods to policy phrases. Our empirical evaluation on 477 top Android apps discovered 341 potential privacy policy violations.

A fundamental problem in the specification of regulatory privacy policies such as the Health Insurance Portability and Accountability Act (HIPAA) in a computer system is to state the policies precisely, consistent with their high-level intuition. In this paper, we propose UML sequence diagrams as a practical means to graphically express privacy policies. A graphical representation allows decision-makers such as application domain experts and security architects to easily verify and confirm the expected behavior. Once intuitively confirmed, our work in this article introduces an algorithmic approach to formalizing the semantics of sequence diagrams in terms of linear temporal logic (LTL) templates. In all the templates, different semantic aspects are expressed as separate, yet simple LTL formulas that can be composed to define the complex semantics of sequence diagrams. The formalization enables us to leverage the analytical powers of automated decision procedures for LTL formulas to determine if a collection of sequence diagrams is consistent, independent, etc. and also to verify if a system design conforms to the privacy policies. We evaluate our approach by modeling and analyzing a substantial subset of HIPAA rules using sequence diagrams.

Security requirements patterns represent reusable security practices that software engineers can apply to improve security in their system. Reusing best practices that others have employed could have a number of benefits, such as decreasing the time spent in the requirements elicitation process or improving the quality of the product by reducing product failure risk. Pattern selection can be difficult due to the diversity of applicable patterns from which an analyst has to choose. The challenge is that identifying the most appropriate pattern for a situation can be cumbersome and time-consuming. We propose a new method that combines an inquiry-cycle based approach with the feature diagram notation to review only relevant patterns and quickly select the most appropriate patterns for the situation. Similar to patterns themselves, our approach captures expert knowledge to relate patterns based on decisions made by the pattern user. The resulting pattern hierarchies allow users to be guided through these decisions by questions, which introduce related patterns in order to help the pattern user select the most appropriate patterns for their situation, thus resulting in better requirement generation. We evaluate our approach using access control patterns in a pattern user study.