Dark Reading is part of the Informa Tech Division of Informa PLC

This site is operated by a business or businesses owned by Informa PLC and all copyright resides with them.Informa PLC's registered office is 5 Howick Place, London SW1P 1WG. Registered in England and Wales. Number 8860726.

Operational Security //

Machine Learning

11/22/2018
09:05 AM
Larry Loeb
Larry Loeb
Larry Loeb
50%
50%

Perceptual Ad Blockers Have Security Flaws, Too

Blocking ads is more than stopping annoying pop-ups. There's a security component as well. However, a crop of perceptual ad blockers that use machine learning have their own flaws and shortcomings.

Users want to be able to block the ads that show up on websites for more reasons than not wanting to be annoyed by the content. They also wish to protect themselves from the malicious content that may be embedded within those ads.

Typical ad blockers currently in use are based on "filter lists," which are specifics that tell the ad blockers exactly what type of content to block.

However, there are problems with this approach, notably keeping that list current and relevant. This led researchers at Princeton University to develop what they called a perceptual ad blocker over the course of 18 months.

Essentially, it looks at the web page, and deletes ads based on what it sees.

It was hoped that this kind of approach would stop the "arms race" of ad generators making minor tweaks in their ads to stop the blocking actions, which was then followed by ad blockers responding with new lists, and then ad generators making further tweaks.

Perceptual blockers are designed to pick up on structural ad cues like a "Sponsored" link or a close button inside of a pop-up ads, as well as any legally required signifiers an ad must show. They will use these structural elements to make the blocking decision.

But security researchers from Stanford University and CISPA Helmholtz Center for Information Security have found some conceptual and practical problems with this approach. Their research paper finds that they have found ways that such a blocker not only can be definitively defeated, but may pose security problems.

As the researchers note:

We show that perceptual ad-blocking engenders a new arms race that likely disfavors ad-blockers. Unexpectedly, perceptual ad-blocking can also introduce new vulnerabilities that let an attacker bypass web security boundaries and mount DDoS attacks.

The researchers looked at two perceptual ad blockers.

The first is named Perceptual Ad Highlighter and the other is called Sentinel. Both use neural networks to do the training needed for the blockers to learn what an ad looks like. From this training, the blockers create a model of what they expect an ad to look like. This model is then used for the "block or leave alone" decision.

To attack the perceptual blockers, the researchers first created ads that were designed to fool both of the blockers. They made changes to the form of the ads that were imperceptible to a human, but capable of fooling a machine learning algorithm.

One change they made was to alter the AdChoices logo commonly found in ads. They also found that a transparent mask laid over a website could do the trick as well. The researchers also looked at other possible attacks.

However, in the end, they found that current ways of developing perceptual ad blockers were not going to survive a determined adversary. The researchers feel that the methods that are currently in use to train such blockers are easily hoodwinked.

As the researchers summarized:

The overarching goal of this analysis has been to highlight and raise awareness of the fundamental vulnerabilities that perceptual ad-blockers inherit from existing vision classifiers. As long as robust defenses to adversarial examples elude us, perceptual ad-blockers will be dragged into a new arms race in which they start from a precariously disadvantaged position -- given the stringent threat model that they must survive.

Related posts:

— Larry Loeb has written for many of the last century's major "dead tree" computer magazines, having been, among other things, a consulting editor for BYTE magazine and senior editor for the launch of WebWeek.

Comment  | 
Print  | 
More Insights
Comments
Newest First  |  Oldest First  |  Threaded View
Manchester United Suffers Cyberattack
Dark Reading Staff 11/23/2020
As 'Anywhere Work' Evolves, Security Will Be Key Challenge
Robert Lemos, Contributing Writer,  11/23/2020
Cloud Security Startup Lightspin Emerges From Stealth
Kelly Sheridan, Staff Editor, Dark Reading,  11/24/2020
Register for Dark Reading Newsletters
White Papers
Video
Cartoon Contest
Write a Caption, Win an Amazon Gift Card! Click Here
Latest Comment: This comment is waiting for review by our moderators.
Current Issue
2021 Top Enterprise IT Trends
We've identified the key trends that are poised to impact the IT landscape in 2021. Find out why they're important and how they will affect you today!
Flash Poll
Twitter Feed
Dark Reading - Bug Report
Bug Report
Enterprise Vulnerabilities
From DHS/US-CERT's National Vulnerability Database
CVE-2019-20934
PUBLISHED: 2020-11-28
An issue was discovered in the Linux kernel before 5.2.6. On NUMA systems, the Linux fair scheduler has a use-after-free in show_numa_stats() because NUMA fault statistics are inappropriately freed, aka CID-16d51a590a8c.
CVE-2020-29368
PUBLISHED: 2020-11-28
An issue was discovered in __split_huge_pmd in mm/huge_memory.c in the Linux kernel before 5.7.5. The copy-on-write implementation can grant unintended write access because of a race condition in a THP mapcount check, aka CID-c444eb564fb1.
CVE-2020-29369
PUBLISHED: 2020-11-28
An issue was discovered in mm/mmap.c in the Linux kernel before 5.7.11. There is a race condition between certain expand functions (expand_downwards and expand_upwards) and page-table free operations from an munmap call, aka CID-246c320a8cfe.
CVE-2020-29370
PUBLISHED: 2020-11-28
An issue was discovered in kmem_cache_alloc_bulk in mm/slub.c in the Linux kernel before 5.5.11. The slowpath lacks the required TID increment, aka CID-fd4d9c7d0c71.
CVE-2020-29371
PUBLISHED: 2020-11-28
An issue was discovered in romfs_dev_read in fs/romfs/storage.c in the Linux kernel before 5.8.4. Uninitialized memory leaks to userspace, aka CID-bcf85fcedfdd.