1 of 21

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

Published at: MDPI Journal of Imaging [Journal]

Mohan Krishna Sunkara

PhD Candidate

Department of Computer Science

Old Dominion University

Norfolk, VA

April 26, 2024 @mk344567, @WebSciDL

Special Issue in Image and Video Processing for Blind and Visually Impaired

2 of 21

Introduction - Deceptive Ads' Impact on Blind Users

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

@mk344567, @WebSciDL

Complex HTML document object models (DOMs) and the absence of textual alternatives for visual components make web pages difficult to navigate for blind users.

Blind users are easily deceived by advertisements that seamlessly integrate with the web page content.

Many ad blockers fail to remove internal advertisements placed by websites.

Many websites are blocking ad blockers

[7]

Visual impairment affects millions worldwide, with 2.2 billion having vision impairment and 36 million being blind.

Blind users are often deceived by ads that seamlessly blend in with the content of the web page. These advertisements can be particularly misleading, making it challenging for blind users to differentiate between genuine content and sponsored ads. Ads are visually rich and often deceive blind users, impacting their browsing experience. Only a small fraction use ad blockers due to installation and configuration challenges.

Furthermore, many websites have started blocking ad blockers, adding another layer of difficulty for blind users. Even when ad blockers are used, they often fail to prevent internal advertisements placed by websites, further complicating the browsing experience for blind users.

3 of 21

How do Blind People Interact with the Web?

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Read out Content on the screen and enable content navigation via keyboard shortcuts

@mk344567, @WebSciDL

[1]

[2]

[3]

[6]

4 of 21

Dark Patterns and Deceptive Content

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Dark patterns manipulate user behavior, and deceptive online ads pose challenges, especially for blind users relying on screen readers.

Lack of visual context in ads can lead to unintentional interactions for blind users.

@mk344567, @WebSciDL

[4]

[5]

[7]

Dark patterns, refer to deceptive user interface designs that manipulate user behavior, often leading to unintended actions such as purchases or sign-ups. These patterns have become prevalent across various digital platforms, leveraging cognitive biases and hiding essential information to deceive users.

While academic research has explored dark patterns extensively, there has been relatively less focus on deceptive online ads. However, most of this research has been conducted from the perspective of sighted interaction, overlooking the unique challenges faced by blind users. As ads often contain visual elements, blind users relying on screen readers may not receive proper auditory feedback to distinguish between ads and regular content.

This lack of visual context increases the risk of blind users inadvertently interacting with deceptive ads. Therefore, there is a need for focused analysis to understand the challenges blind users face with ads, which we address in our work.

5 of 21

Ad Blocking and Ad Detection Efforts

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Efforts to improve ad accessibility are emerging but face challenges in implementation and adoption, highlighting the importance of detecting and addressing deceptive ads for blind users.

@mk344567, @WebSciDL

[4]

[5]

[7]

Various ad detection browser add-ons like AdGuard, AdBlock, and AdBlock Plus are available to help users avoid ads on web pages. These extensions work by blocking communications to ad servers and hiding ads from the HTML DOM. However, they may not eliminate internal promotions or ads, which are often deceptive and contribute significantly to website income.

Despite the availability of ad blockers and detection algorithms, many websites discourage ad blocking due to revenue loss. Some websites employ techniques to detect ad blockers and may deny access to content or services upon detection.

Efforts are also underway to advocate for the inclusion of accessibility guidelines in online advertising practices. However, these efforts are still in their early stages and may take time to yield positive outcomes. Additionally, creators of deceptive ads may not follow accessibility guidelines, highlighting the need for approaches that detect and communicate the presence of deceptive ads to screen-reader users

6 of 21

Interview study: Understanding Blind Users Perspectives on Deceptive Content

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Participant recruitment: 18 blind participants, aged 22 to 66, were recruited. They were familiar with screen readers and web browsing, with daily web usage.
Inclusion/exclusion criteria: Proficient in English, solely reliant on screen readers for web access. Excluded mildly visually impaired individuals and children under 18.
Interview setup: Remote Zoom interviews focused on the impact of ads and promotions on browsing experience.
Data collection: Recorded sessions with consent, no identifiable information retained. Encouraged live demonstrations of interaction issues.
Data analysis: Interview data analyzed using open and axial coding techniques to identify recurring insights.

@mk344567, @WebSciDL

We recruited 18 blind participants, comprising 11 females and 7 males, for our study. Their ages ranged from 22 to 66, with an average age of 41.7. All participants were well-versed in screen readers and web browsing and communicated fluently in English.

The interviews were conducted remotely via Zoom, utilizing a semi-structured format. We focused on topics such as the impact of advertisements and promotions on browsing experiences, participants' browsing strategies, and their interaction behavior regarding ads. Participants were encouraged to share their screens and demonstrate their browsing behaviors during the sessions.

Each interview lasted approximately 45 minutes and was recorded with participants' consent. The collected data underwent rigorous analysis using open coding and axial coding techniques to identify recurring insights and themes.

7 of 21

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Interview Study Participant Demographics

@mk344567, @WebSciDL

8 of 21

Interview Study: Findings

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Ad Blocker Challenges: Most blind users avoid ad blockers (15 out of 18) due to installation and configuration difficulties, causing accessibility issues and website access problems.
Increased Web Task Difficulty: Ads prolong web activities for blind users due to screen reader's linear navigation, making navigation tedious.
Encountered Ad Types: Common types include e-commerce deals, memberships, and native promotions, with the latter often deceptive.
Importance of Ad Location: Participants often accidentally select ads strategically placed to blend with content, leading to wasted effort.
Screen Reader Limitations: Screen readers inadequately describe ads, causing discrepancies in sighted and blind users' perceptions due to missing visual information.
Cautious Interaction and Sighted Assistance: Blind users are cautious with ads and often avoid interaction without sighted assistance for safety. Upon accidental selection, they promptly close the browser and delete downloaded files.

@mk344567, @WebSciDL

Our qualitative analysis of the interview responses from blind screen-reader users shed light on several key themes related to their experiences with online ads:

Ad Blocker Challenges: Most participants avoided using ad blockers due to difficulties in installation and configuration, leading to accessibility issues and website access problems. Some even reported getting stuck on websites due to inaccessible notification pop-ups.

Increased Web Task Difficulty: Participants found web activities more time-consuming due to ads, attributing this to the linear navigation nature of screen readers, which requires them to listen through content to identify and skip ads.

Encountered Ad Types: Participants commonly encountered e-commerce deals, membership sign-ups, and native promotions. While some ads were considered harmless, others were distracting or deceptive.

Importance of Ad Location: The location of ads significantly impacted browsing experiences, with participants often being deceived by ads blending with surrounding content.

Screen Reader Limitations: Screen readers inadequately described ads, leading to discrepancies between sighted and blind users' perceptions.

Cautious Interaction and Sighted Assistance: Participants were cautious with ads and often avoided interaction without sighted assistance for safety. Upon accidental selection, they promptly closed the browser and deleted downloaded files.

9 of 21

Deceptive Content Detection Algorithm

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

@mk344567, @WebSciDL

Our algorithm begins by traversing the DOM tree using depth-first search to identify potential deceptive ad candidates based on these specific node types.

Once potential candidates are identified, we extract information from these candidate tags and derive both handcrafted and automatic features.

Handcrafted features capture contextual information like web page summaries, similarity scores between summaries and candidate texts, and candidate URL features. These features are binary, represented as either 0 or 1.

Automatic features, derived from images and text within the candidate DOM sub-tree, are represented as numerical floating point values.

The next step involves classifying candidates as either deceptive or non-deceptive. For this, we trained a custom multi-modal classifier that combines both handcrafted and automatically extracted features.

The classifier consists of a transformer encoder with an appended LSTM layer, which takes in concatenated features.

Finally, the algorithm determines the deceptive nature of ads based on the output of the multi-modal classifier.

10 of 21

Algorithm Workflow

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

@mk344567, @WebSciDL

The model utilizes three sets of features for classification. The first set is derived from the ad text, including text extracted from images using Tesseract OCR and processed with a pre-trained BERT model.

The second set of features comes from image elements in the ad, extracted using a pre-trained Vision Transformer model.

Additionally, we incorporate handcrafted features based on patterns observed in 50 sample web pages, such as the absence of ARIA attributes in deceptive ads and contextual information like web page summaries.

Before feature extraction, we pre-process the input data by cleaning textual data, resizing images, and normalizing pixel values.

These features are then concatenated and passed through a neural pipeline consisting of 12 transformer encoder layers, an LSTM layer, and a dense layer.

Each encoder layer employs multi-head self-attention mechanisms and feed-forward neural networks, while the LSTM layer includes 128 hidden units with a tanh activation function.

The final dense layer utilizes the softmax activation function to determine the class probabilities, with the output class being the one with the highest probability.

This architecture enables the model to effectively classify ads as either deceptive or non-deceptive based on the input features.

11 of 21

Hand Crafted Features

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

@mk344567, @WebSciDL

Web page summary: This attribute uses the summary of the web page to determine if the ad is deceptive. This suggests that the content of the ad itself may not be enough to determine if it's deceptive, and that other information from the landing page is also important.

Context similarity: This attribute looks for similarity between the ad text and the summary of the web page. This suggests that deceptive ads may try to mislead users by not matching the content of the ad to the content of the landing page.

Readability score: This attribute considers the Flesch Reading Ease score of the ad text. This score measures how easy it is to read a piece of text. Deceptive ads may tend to use language that is difficult to understand.

URL security: This attribute checks if the ad's URL uses HTTPS, which is a secure protocol. Deceptive ads may tend to use non-secure URLs.

URL with IPv4 address: The model checks if the ad's URL contains an IPv4 address. Deceptive ads may tend to avoid URLs with IPv4 addresses.

URL active/inactive: This attribute checks if the URL linked in the ad is still active. Deceptive ads may tend to use URLs that are only active for a short period of time.

Number of URL redirects: The model checks how many times a user is redirected when they click on the URL in the ad. Deceptive ads may tend to use URLs that include multiple redirects.

12 of 21

Deceptive Content-Detection Dataset

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Web Page Collection: 500 pages from diverse domains like travel, news, blogs, and e-commerce sites were collected for model training.

Web Page Annotation: Deceptive ads were manually annotated, primarily within specific DOM nodes. Custom data attribute data-deceptive="true/false" was added to these nodes.

Dataset Construction: Deceptive and non-deceptive ads were extracted based on the data-deceptive attribute. A supervised dataset (X, y) was created, ensuring balance with 1200 examples.

Availability: The annotated dataset and code for dataset construction are publicly accessible on GitHub and Google Drive.

@mk344567, @WebSciDL

We collected 500 web pages from diverse domains such as travel, news, blogs, and e-commerce sites to train our classification model. This ensured a comprehensive representation of online content with prevalent advertisements.

Next, we manually annotated deceptive ads found within specific DOM nodes like <div>, <iframe>, <script>, <a>, <img>, <span>, and <ins>. We added a custom data attribute, data-deceptive="true/false," to these DOM nodes for annotation.

From these annotated web pages, we extracted both deceptive and non-deceptive ads based on the presence of the data-deceptive attribute. This allowed us to construct a supervised dataset (X, y) where X represents features for each ad, and y indicates the target label (1 for deceptive, 0 for non-deceptive). We ensured balance by sampling 1200 examples with an equal number of deceptive and non-deceptive ads.

Our annotated dataset, along with the code for dataset construction and feature extraction, is publicly available on GitHub and Google Drive. This promotes transparency and reproducibility in research

13 of 21

Model Training

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Hardware: Utilized NVIDIA V100 GPU with 128 GB memory per node for training.

Model: Built using Adam optimizer, binary cross-entropy loss function.

Training: Conducted over 25 epochs with 0.1 validation split, 500 steps per epoch.

Callbacks: Employed ReduceLROnPlateau, EarlyStopping to optimize training.

Optimization: Set parameters for callbacks to ensure efficient convergence.

@mk344567, @WebSciDL

For our model training, we utilized an NVIDIA V100 GPU with 128 GB of memory per node, providing ample resources for efficient processing.

We built the model using the Adam optimizer and a binary cross-entropy loss function, which proved effective for our classification task. The learning rate was dynamically adjusted based on the model's performance on the validation data.

Training was conducted over 25 epochs with a validation split of 0.1, consisting of 500 steps per epoch. This ensured thorough coverage of the dataset while monitoring performance.

To optimize the training dynamics, we implemented callback functions such as ReduceLROnPlateau and EarlyStopping. ReduceLROnPlateau dynamically adjusted the learning rate based on validation loss changes, while EarlyStopping prevented overfitting by halting training when necessary.

Strategic parameter settings for these callbacks ensured efficient convergence towards optimal model parameters, ultimately improving performance and enhancing the model training process

14 of 21

30% Surge in Precision, Recall, F1, and Accuracy

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Precision, Recall, F1 Score, Accuracy increased by approximately 30 percent

@mk344567, @WebSciDL

As part of our evaluation process, we allocated 20% of the supervised dataset for testing the model's performance. This test data comprised examples from various websites across different domains, ensuring a comprehensive assessment.

Using standard classification measures such as precision, recall, accuracy, and F1 score, we evaluated several model configurations. These configurations included ResNet50 and Vision Transformer combined with BERT, both with and without additional features.

Our findings indicate that incorporating extra features significantly improves model performance. The proposed model, equipped with these additional features, consistently outperformed other configurations, achieving high scores across all metrics, including precision (0.865), recall (0.862), accuracy (0.863), and F1 score (0.863).

Even the baseline model demonstrated strong performance, with scores ranging from 0.794 to 0.800. However, the ResNet50 + BERT and Vision Transformer + BERT combinations showed varying performance levels, with notable improvements observed when additional features were integrated.

Furthermore, we conducted an ablation study to assess the impact of different handcrafted features on model performance. The results, detailed in Table 4, provide insights into the contributions of these features to overall model effectiveness.

15 of 21

Ablation study

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

handcrafted features: α = “ARIA Attribute”, β = “Readability Score”, γ = “Is URL Secure”, δ = “URL

Host Name”, ε = “URL Active/Inactive”, and ζ = “Number of URL Redirections”.

@mk344567, @WebSciDL

Our evaluation included an ablation study focusing on handcrafted features, such as ARIA attribute, readability score, URL security, URL host name, URL activity, and number of URL redirects. The baseline model, comprising automatic features along with web page summary and context similarity, served as a reference point. Additional contextual features generally improved model accuracy from 0.79 (baseline) to 0.86 (with all features).

While overall improvements were observed with extra features, some exceptions occurred. For instance, the addition of the number of URL redirects resulted in a slight drop in the F1 score from 0.865 to 0.863.

We further evaluated the algorithm's performance 'in-the-wild' by testing it on 20 randomly selected web pages from various domains. These pages represented diverse online contexts, including travel, hotel, news, and blogs. The algorithm achieved an impressive F1 score of 0.88, demonstrating its effectiveness in identifying deceptive ads across different websites.

16 of 21

Discussion

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Limitations: The algorithm's applicability may be limited to specific dark patterns. Evaluation is constrained by a small sample size focused on English-language web pages. Represents an initial step in addressing usability issues.

Expansion Plans: Aim to expand datasets for validation, build annotated datasets for various dark patterns, and experiment with alternative deep neural architectures and pre-trained models.

Assistive Technology Development: Develop innovative browser extensions that replace deceptive ads with screen reader-friendly content, thereby enhancing accessibility and user safety for blind screen-reader users.

Societal Impact: Contributes to improving web accessibility for visually impaired individuals, enabling informed interactions with online ads, and fostering a more inclusive digital environment.

@mk344567, @WebSciDL

Our study focused solely on detecting deceptive ads, which may not extend to identifying other types of dark patterns like privacy zuckering, roach motel, confirmshaming, and bait-and-switch. Detecting these patterns might necessitate the development of new algorithms tailored to each type.

Additionally, our algorithm's evaluation was constrained by a small sample due to the labor-intensive manual annotation process. We acknowledge this limitation and are actively working on expanding our annotated dataset to ensure a more comprehensive evaluation of our algorithm's performance.

Furthermore, while our algorithm represents a crucial first step in addressing usability issues for blind screen-reader users interacting with web page ads, there's still room for improvement. Our future research will focus on designing intelligent interactive systems that leverage our algorithm to enhance user experience.

Lastly, our algorithm was tested exclusively on English-language web pages. While its design is adaptable to other languages, testing on non-English datasets is essential to assess its performance across different languages and ensure its effectiveness in diverse linguistic contexts

17 of 21

Discussion

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Limitations: The algorithm's applicability may be limited to specific dark patterns. Evaluation is constrained by a small sample size focused on English-language web pages. Represents an initial step in addressing usability issues.

Expansion Plans: Aim to expand datasets for validation, build annotated datasets for various dark patterns, and experiment with alternative deep neural architectures and pre-trained models.

Assistive Technology Development: Develop innovative browser extensions that replace deceptive ads with screen reader-friendly content, thereby enhancing accessibility and user safety for blind screen-reader users.

Societal Impact: Contributes to improving web accessibility for visually impaired individuals, enabling informed interactions with online ads, and fostering a more inclusive digital environment.

@mk344567, @WebSciDL

In our future work, we plan to enhance our research in several key areas. Firstly, we aim to expand the size of our datasets to further validate the effectiveness of our classification model and algorithm. This expansion will provide more robust evidence of our approach's reliability.

Additionally, we intend to broaden the scope of our algorithm by creating annotated datasets and training classifiers to detect various types of dark patterns beyond deceptive ads. These may include bait-and-switch, roach motel, and other emerging patterns.

Furthermore, we will explore the integration of different deep neural architectures, such as PaLI and ViLBERT, known for their effectiveness in multi-modal classification tasks. Experimenting with these architectures will enable us to leverage their capabilities and potentially enhance the performance of our algorithm.

Lastly, we plan to experiment with various pre-trained models for automatic image and text feature extraction, including BERT-large-uncased and Swin Transformers. By incorporating these models and optimizing their performance-enhancing methods, we aim to further improve the accuracy and efficiency of our algorithm."

18 of 21

Discussion

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Limitations: The algorithm's applicability may be limited to specific dark patterns. Evaluation is constrained by a small sample size focused on English-language web pages. Represents an initial step in addressing usability issues.

Expansion Plans: Aim to expand datasets for validation, build annotated datasets for various dark patterns, and experiment with alternative deep neural architectures and pre-trained models.

Assistive Technology Development: Develop innovative browser extensions that replace deceptive ads with screen reader-friendly content, thereby enhancing accessibility and user safety for blind screen-reader users.

Societal Impact: Contributes to improving web accessibility for visually impaired individuals, enabling informed interactions with online ads, and fostering a more inclusive digital environment.

@mk344567, @WebSciDL

Our algorithm lays the groundwork for innovative assistive technologies that can significantly improve the browsing experience of blind users, particularly in dealing with online ads. One such technology involves replacing deceptive ads with screen reader-friendly content. By identifying ad elements on a webpage, extracting relevant information, and injecting descriptive text alternatives, visually impaired users can better understand the content and make informed decisions while browsing, enhancing both accessibility and user safety.

Another potential technology is the automatic generation and injection of context-relevant textual summaries for ads into the corresponding webpage DOM node. This allows screen readers to provide more detailed information to users, aiding in their decision-making process regarding ad interaction.

Additionally, we propose the implementation of informative skip links before deceptive ads in the webpage DOM. This enables screen-reader users to bypass ad content altogether, ensuring they do not inadvertently engage with ads without needing to filter them out manually.

In our future research, we aim to explore and develop these ideas further, ultimately creating assistive technologies that empower blind users to navigate online content with greater ease and confidence.

19 of 21

Discussion

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Limitations: The algorithm's applicability may be limited to specific dark patterns. Evaluation is constrained by a small sample size focused on English-language web pages. Represents an initial step in addressing usability issues.

Expansion Plans: Aim to expand datasets for validation, build annotated datasets for various dark patterns, and experiment with alternative deep neural architectures and pre-trained models.

Assistive Technology Development: Develop innovative browser extensions that replace deceptive ads with screen reader-friendly content, thereby enhancing accessibility and user safety for blind screen-reader users.

Societal Impact: Contributes to improving web accessibility for visually impaired individuals, enabling informed interactions with online ads, and fostering a more inclusive digital environment.

@mk344567, @WebSciDL

While accessibility aims to provide equal content access for all, it doesn't always guarantee optimal interaction usability, particularly for users reliant on screen readers.

This research addresses a specific accessibility challenge related to online ads and promotions. By facilitating more informed interactions with these elements, blind screen-reader users can benefit in various ways. They can take advantage of online opportunities and deals, avoid unintended transactions, prevent accidental downloads, develop efficient web navigation strategies, and enhance overall web activities with reduced security and privacy concerns.

Our algorithm represents a significant first step towards addressing these challenges by automatically identifying deceptive ads on web pages. Future efforts can build upon this work to create more informative interactions with online ads and promotions. These outcomes will not only enhance the personal experiences of blind users but also contribute to a more inclusive and fair digital environment, allowing blind individuals to participate more seamlessly in online interactions.

20 of 21

Summary

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

Knowledge Gap: Explored impact of deceptive content on blind screen-reader users.
Interview Study: Engaged 18 blind participants, unveiling susceptibility to deceptive ads.
Detection Algorithm: Created an effective algorithm for identifying deceptive ads.
Performance: Achieved high F1 scores: 0.86 on a representative dataset, 0.88 on an "in-the-wild" evaluation of web pages.
Future Focus: Algorithms for other kinds of dark patterns.

Dataset: https://drive.google.com/drive/folders/1UemGmaBLcZ9SWHY0m28Krnc6eyn6OdlV

Webpage: https://www.cs.odu.edu/~cs_msunk001/

Twitter: https://twitter.com/mk344567�Email: msunk001@odu.edu

@mk344567, @WebSciDL

In this paper, we identified a knowledge gap concerning the impact of web content, such as ads and promotions, on blind screen-reader users' interaction experience. Through an interview study involving 18 blind participants, we discovered that these users are highly vulnerable to deceptive advertisements and promotions.

To tackle this issue, we developed a novel algorithm capable of identifying contextually deceptive ads on web pages. This algorithm, powered by a custom classification model, utilizes a combination of hand-crafted and automatically extracted features. Upon testing, the model achieved an impressive F1 score of 0.86 on a representative dataset and 0.88 during an "in-the-wild" evaluation on 20 randomly selected web pages.

Given its strong performance, we believe our algorithm will lay the groundwork for future solutions aimed at enhancing the usability of non-visual ad interactions

21 of 21

References:

Detecting Deceptive Dark-Pattern Web Advertisements for

Blind Screen-Reader Users

WS-DL 2024 Research Expo

NV Access. "About NVDA." Accessed on https://www.nvaccess.org/about-nvda/.
Apple Support. "VoiceOver User Guide for Mac." Accessed on https://support.apple.com/guide/voiceover/welcome/mac.
Freedom Scientific. "JAWS Screen Reader." Accessed on https://www.freedomscientific.com/products/software/jaws/.
Google Play. "Ad Blocker for Android." Accessed on https://play.google.com/store/apps/details?id=jp.snowlife01.android.ad_blocker&hl=en_US.
Facebook. "Adblock by Adblock." Accessed on https://www.facebook.com/heres.adblock/.
Priceline. Accessed on https://www.priceline.com/.
KAYAK. Accessed on https://www.kayak.com/.

@mk344567, @WebSciDL

Given its strong performance, we believe our algorithm will lay the groundwork for future solutions aimed at enhancing the usability of non-visual ad interactions