Using Salient Object Detection to Identify Manipulative Cookie Banners that Circumvent GDPR

AI summary

Overview

Research area: Web privacy and AI ethics — automated detection of dark patterns (specifically "aesthetic manipulation") in GDPR cookie banners using computer vision. Technical level: Intermediate. The paper combines a legal/regulatory framing with a computer-vision technique (salient object detection) and descriptive statistics; no deep ML background is required to follow the argument, but familiarity with salience models and consent regulation helps. Scope: The paper measures how often GDPR-compliant cookie banners nevertheless use visual design to steer users toward accepting personal-data sharing, across 2,579 EU and non-EU websites visited from both EU and US IP addresses.

What This Paper Is About

Websites are legally required by the GDPR to offer users a real choice about personal-data collection through cookie banners, but many design those banners so that the "accept" button draws more attention than the "reject" button — a dark pattern called aesthetic manipulation. The paper's core problem is that prior studies measured this pattern indirectly, mostly by checking whether the accept button was highlighted, which misses subtler tricks such as where a button is placed or how large it is. The goal is to measure salience directly with a computer-vision salience model and thereby produce a more accurate, more current estimate of how widespread aesthetic manipulation is, and how it varies with website and visitor location.

Key Contributions

1. A new measurement method for aesthetic manipulation. The authors are the first to apply a salient object detection (SOD) model, DeepRare, directly to cookie banner screenshots to compute how salient each button is, rather than relying on a single predefined design cue such as button highlighting.
2. A large, public, two-location cookie banner dataset. They collected landing-page screenshots from 2,579 websites — popular and randomly sampled, EU and non-EU — crawled between December 16 and 23, 2024 from VPN IP addresses in Paris and New York, and manually classified banners into 17 design categories. The data and code are released at github.com/rag24/aesthetic-manipulation-SOD.
3. Evidence that prior work undercounted aesthetic manipulation. Using DeepRare, they find aesthetic manipulation on 38.0% of GDPR-compliant banners, whereas methods from prior work would have flagged only 27.0% of the same banners.
4. New forms of aesthetic manipulation and location effects. They report that displaying a button as a link, button placement, and button size are all significantly correlated with button salience, and that EU websites are roughly 48.3% more likely to use aesthetic manipulation than non-EU websites, with 13.9% of EU websites changing their banner design depending on whether the visitor arrives from the US.

Main Findings

• Compliance is far from universal. Of the relevant websites, 45% have fully compliant banners, yet aesthetic manipulation appears on 38% of those compliant banners.
• Compliance depends sharply on location. Fully compliant banners are 44.9% for EU sites visited from an EU IP, 40.9% for EU sites visited from a US IP, 48.4% for non-EU sites visited from an EU IP, and only 13.5% for non-EU sites visited from a US IP. Non-compliant banners rise to 63.0% in that last case.
• EU websites personalize banners by visitor location, despite GDPR applying either way. Among the 1,003 EU websites reachable from both EU and US IP addresses, 139 (13.9%) show a different design to US visitors. Of those, 115 (82.7%) remove the banner entirely and 14 (10.1%) swap a compliant banner for a less compliant one; of the 115 that remove the banner, 64 (55.7%) had shown a fully compliant banner to EU visitors.
• Cookie paywalls are an EU-only phenomenon in this dataset. They appear on 6.3% of EU websites for EU visitors and 5.4% for US visitors, and are nonexistent on non-EU websites.
• Aesthetic manipulation is more common than previously reported. 38.0% of compliant banners versus 27.0% detectable with prior approaches.
• New salience-driving design factors. Button placement, button size, and rendering a button as a link are significantly correlated with salience, and the authors argue these should be treated as forms of aesthetic manipulation in addition to highlighting.
• Opt-out rates have improved over time. Reject options are available on 94.9% of banners on EU websites seen from the EU and 88.9% on non-EU websites seen from the EU — the highest percentages reported among the studies compared.
• Accept-and-reject buttons on the first page have become much more common. 60.6% for EU websites (EU visitor) and 68.5% for non-EU websites (EU visitor), higher than any prior dataset compared; the authors link this to CNIL fines against Meta and Google in December 2021.
• Preselected consent options are now rare. Under 2% of banners, down from 46.5%–56.2% in 2019, following the Planet 49 ruling.
• Fewer double banners. Roughly 1% of websites show two cookie banners, versus 7.9% in a 2022 study; the authors state the reason for the decrease is not clear.

Methodology in Plain English

The authors built a target list of 3,000 websites from Tranco rankings: the top 1,000 global sites plus 1,000 randomly sampled global sites, and the top 500 EU-ccTLD sites plus 500 randomly sampled EU sites. They note that 44 sites appeared on both lists, and they removed 377 that were not reachable on standard HTTP/HTTPS ports, leaving 2,579 sites.

A Selenium-based crawler screenshotted each landing page. It tried HTTPS first with a 30-second timeout, fell back to HTTP, then retried at 60 seconds, and finally resorted to manual visiting. Each site was visited twice through VPNs — once from Paris (EU) and once from New York, chosen because the New York Privacy Act has not been enacted. The crawler captured 2,490 landing pages from Paris and 2,492 from New York.

Every screenshot was manually labeled by a rater as unreachable, banner-free, or one of 17 banner design categories (for example "Notice," "Paywall," "Full," "Choices," "Manage," "Corner Reject," "Preselected," "Ambiguous," "Two Banners"). Images were labeled twice, disagreements were resolved by revisiting the site, and two additional PhD students labeled 200 randomly sampled sites for quality control, yielding a Krippendorff's alpha of 0.94.

For the salience analysis, they restricted attention to banners with multiple buttons (the "Full," "Full choices," "Choices," "Manage," "Full-Manage," and "Corner Reject" categories). Each banner element — the banner itself, accept button, reject button, manage-settings button — was outlined with bounding boxes using Label Studio. Screenshots were then fed to DeepRare, which outputs a pixel-level salience score between 0 and 255 without requiring additional training. To reduce noise, the authors perturbed each image slightly (shifting hues, flipping it vertically or horizontally) and averaged salience scores across the perturbed versions. Button-level salience was computed as the sum of the normalized average and normalized maximum pixel salience inside each button's bounding box. Drawing on Weber's Law and reported just-noticeable differences ranging from 1.6% for contrast to 6.8% for angular size, they set a conservative 7% threshold: aesthetic manipulation is counted only when the accept button is at least 7% more salient than the other buttons.

As an illustration, on washingtonpost.com the average salience scores were 161.6 for accept, 160.9 for reject, and 84.5 for manage settings (maximums 255, 252, and 169), showing the manage-settings button was clearly less salient while accept and reject were nearly tied.

Why This Matters

Impact on research. The paper shows that operationalizing a dark pattern through a single design cue (highlighting) systematically understates how common that pattern is — 27.0% versus 38.0% on the same banners. It introduces computer vision, specifically salience modeling without task-specific training, as a measurement instrument for privacy research, and it is the second study to test all four combinations of EU/non-EU website and EU/US visitor location, and the first to do so for aesthetic manipulation.

Real-world applications:

• Regulators such as national data protection authorities could use salience scoring as an objective test when evaluating whether a banner design constitutes a dark pattern.
• Browser extension and browser-vendor developers could flag or automatically neutralize high-salience accept buttons on the pages users visit.
• Website compliance and legal teams could audit their own banners before publication rather than after a complaint.
• Privacy advocacy organizations could scale dark-pattern audits beyond manual review of a few hundred sites.

Industry relevance. The finding that EU websites are roughly 48.3% more likely to use aesthetic manipulation, and that 13.9% of them show different banners to US versus EU visitors, is directly relevant to ad-tech and publishing companies operating in regulated markets. The paper frames aesthetic manipulation as a way to comply with the letter of GDPR while still nudging users toward data sharing — a legal-but-against-the-spirit strategy, alongside cookie paywalls, that firms can be expected to keep pursuing.

Future Directions

• Extend salience measurement beyond landing-page screenshots, since banner elements behind settings pages or multi-step flows are not captured by a single image.
• Establish a legally grounded, quantitative definition of when a salience gap should be considered a violation, given that the authors could not find unified guidance from EU data protection authorities on many of the 17 banner categories, including "Corner Reject."
• Re-run the study longitudinally to test whether the observed 2024 trends — rising opt-out availability and falling preselect options — continue as new guidance and fines appear, and whether new innovative designs emerge in response.
• Validate DeepRare's salience output against human attention data (for example eye-tracking) rather than relying on Weber's Law-derived thresholds, since the authors note no JND value has been experimentally established for visual salience.

Target Audience

Privacy and web-regulation researchers, HCI and dark-pattern scholars, regulators and policy analysts working on GDPR enforcement, and computer-vision practitioners interested in applying salience models outside traditional vision tasks. The paper is also useful to legal and compliance professionals at companies that deploy cookie banners, and to anyone who wants a current (2024) picture of how cookie banner design has evolved since the GDPR took effect.

Authors’ abstract

The main goal of this paper is to study how often cookie banners that comply with the General Data Protection Regulation (GDPR) contain aesthetic manipulation, a design tactic to draw users' attention to the button that permits personal data sharing. As a byproduct of this goal, we also evaluate how frequently the banners comply with GDPR and the recommendations of national data protection authorities regarding banner designs. We visited 2,579 websites and identified the type of cookie banner implemented. Although 45% of the relevant websites have fully compliant banners, we found aesthetic manipulation on 38% of the compliant banners. Unlike prior studies of aesthetic manipulation, we use a computer vision model for salient object detection to measure how salient (i.e., attention-drawing) each banner element is. This enables the discovery of new types of aesthetic manipulation (e.g., button placement), and leads us to conclude that aesthetic manipulation is more common than previously reported (38% vs 27% of banners). To study the effects of user and/or website location on cookie banner design, we include websites within the European Union (EU), where privacy regulation enforcement is more stringent, and websites outside the EU. We visited websites from IP addresses in the EU and from IP addresses in the United States (US). We find that 13.9% of EU websites change their banner design when the user is from the US, and EU websites are roughly 48.3% more likely to use aesthetic manipulation than non-EU websites, highlighting their innovative responses to privacy regulation.