From UI to Code: Mobile Ads Detection via LLM-Unified Static-Dynamic Analysis

A team of researchers has introduced ADWISE, a novel framework that tackles the complex challenge of detecting mobile advertisements within apps. The core problem is that ads are dynamically fetched from servers and rendered through diverse UIs, making them difficult to locate and trigger automatically. ADWISE formulates this as an LLM-guided, ad-oriented UI exploration problem. It first performs static program analysis to identify the specific UI widgets used to place ads, which the researchers term 'ad widgets'.

It then employs a grounded LLM reasoning loop to navigate the app and trigger these widgets, guided by three complementary signals. These are WTG-based guidance (using a statically built window transition graph), semantic guidance (reasoning about app functionality to prioritize likely user paths), and structural guidance (using retrieval-augmented generation, or RAG, to match current UIs against known ad-heavy layouts). This hybrid static-dynamic approach allows ADWISE to effectively probe complex app interfaces.

The results are significant. In experiments conducted on a benchmark of 100 apps, ADWISE outperformed existing state-of-the-art detection methods by 25.60% in accurately identifying ad widgets. Beyond mere detection, the framework proved highly effective for compliance auditing, uncovering 34.34% more violations across six categories of ad regulations than previous tools. This demonstrates its direct utility for enforcing advertising standards and improving transparency for users and regulators.