A client wants us to screen work machines for pornography. Is it possible?

Question

A long-time client has asked us to help screen their work machines for pornography. They're worried about liability if sensitive materials were found. Their main concerns (for obvious reasons) are video, audio, and image files. If possible, they'd also like to scan text-based documents for inappropriate content. They have a hierarchy of not-for-work content starting with blatantly illegal (I don't have to list details), moving down to obviously offensive, and also including things that may be offensive to some - think lingerie ads, joke cards featuring butt cracks, and anything related to Howie Mandel.

My questions are:

• Is this ethical? I think it is since every employee legally agrees that their work machine belongs to the company and is subject to search. The screenings are not to occur on personal machines brought to work.
• Is it feasible? I've done a lot of image processing/indexing but this seems like a whole new world of complexity.
• Any references to successful techniques for discovering porn?
• Is it appropriate for me to archive the results when something is discovered?

Answer 1

You can do this with 90% Headology, 10% software.

Firstly, quietly scan employees computers, build a database of files and sizes for each employee.

Then leak a memo that all PC's will be scanned for questionable content, i.e. The bosses have a Shazam like program that can identify porn etc.

Then a couple of days later, scan the computers for files and sizes again. Look at any deleted files, are they movie or image files? Then those are the employees you need to keep an eye on.

Routinely scan those employees PC's for images and movies, and manually check them for questionable content.

Answer 2

This is an obvious neural network task. First you need a large training set of images selected by experts in your company.....

A more effective solution is to announce that you will be checking everyones machine for porn NEXT week/month/whatever, then write a simple app that just exercises the disk. I guarantee that the machines will have been cleaned by then.

---

ps - A couple of 'serious' points - you actually don't want to find anything.

If you do find a couple of images in a browser cache then perhaps they hit a bad link or a dodgy popup - remember the teacher fired over whitehouse.com? If you fire/discipline them for this then there is going to be a backlash from workers/union. How would your company work if every click had to be submitted to legal for approval before your workers researched a question or checked a price online?

If you find a stack of porn on a machine how are you going to prove it was put there by that employee? Do you have the sort of security and audit systems that would stand up in court? Do you use (or even know of) an OS where a system admin couldn't put them there and make it look like the user's files?

Plus in my experience the most common locations for porn stashes are on the laptops of CxOs and senior VPs.

It's much better to just arrange for the files to just vanish ahead of time.

Answer 3

This approach of controlling is certainly painful for both employees and IT people. Once anything enters inside the employee machine, there is no sure way of detecting it. You need to stop it entering in the machine at the first place.
The best known practice for this is obviously control over the sites/domain which can be visited. Such list must be available somewhere on the net. Other than this you can also track the number of images, videos the employee has downloaded and from where it has come.
There are chances that the material can come from other than web, like from external hard drive. There could be once a month random scan of the system where you can randomly pick some of the videos and images and check it manually. Not sure how it can be done. But automating of checking the images and videos is certainly out of scope and certainly will be erroneous.
Actually I am not very much with the idea of restricting employees from doing personal stuff. You should trust your employees for this. Your employees should be busy enough in the office so that they don't get any time for this. The more worries are is the employee not doing his/her work right? Or has s/he installed some cracked or hacked software?

Answer 4

There are a number of products in the marketplace that perform "content filtering" of various forms. (A Google search on some obvious terms throws up some obvious candidates.) It is probably a better idea to use one of these products than building a lot of scanning / filtering software from scratch. Another option is to just watch at the borders; e.g. by monitoring external emails and web traffic. Again there are products that do this kind of thing.

While there is no doubt that it is ethical for a company to scan its computers for "bad stuff", this does not mean that there aren't issues.

First issue:

• Determining what is and what is not "objectionable content" is subjective.
• Software for detecting images, videos containing (let us say) "depictions of the naked body" is (AFAIK) likely to be unreliable, resulting in false positives and false negatives.

So ... this means that someone in your customer's organization needs to review the "hits". That costs money.

Second issue: There can be an innocent explanation. The file could have been downloaded by accident, or it could have been planted by a vindictive co-worker. If there is an innocent explanation, the customer's organization needs to be careful what they do / say. (OK this is not really your issue, but you might cop some of the backwash.)

Third issue: Not-withstanding that the company has a right to monitor for objectionable material, a lot of employees will find this distasteful. And if they too far, this will impact on employee morale. Some employees will "walk". Others may take protest action ... e.g. by trying to create lots of false positives. (Again, not really your issue, but ...)

Fourth issue: People can hide objectionable material by encrypting it, by putting it on portable or removable media, etc. People can fake the metadata to make it look like someone else is responsible.

Answer 5

About legal aspects, in France:

The boss owns the computers and the internet connection: He can do whatever pleases him.

BUT, employee privacy cannot be violated. If a directory on the computer is labelled PERSONAL, the boss is not allowed to scan it.

The only way to bypass that is to get elements of evidence that employee stores illegal material and to get a court request a scan of computer (Note that pornography is not illegal in France.)

Answer 6

If the employees agreed that their work machine belongs to the company and is subject to search, then yes, this is legal. For proof, archival of the files would most likely be necessary.

As for how to actually find the material. You could:

1. First and foremost, scan file names for a certain set of words (porn, lesbians, etc.)
2. Scan text documents for the same set of words
3. For images, you could find the average color of the image, and if that color happens to be within a range that most would refer to as 'flesh' colored, then flag the image (someone double checking these flagged images will most likely be necessary). Wouldn't want to report someone for an image that ends up being a family photo from the beach.

If you scan the files as they're entering the computer (e.g. have the program loaded on every work machine and log flagged cases to a central database), then I don't think it would be too obtrusive (other than the blatant distrust the employer clearly has for their employees).

With the video files, I'm not 100% sure. Possibly a similar approach as with the image scanning (choose random frames and scan for a certain level of 'flesh' color).

Scanning audio files seems like it would get into speech recognition, which is a whole 'nother can of worms. Scanning the file name, however, would be easy and could be done as with the documents, images, and video.

Answer 7

As @Ryan said, image analysis can focus on color analysis.

Feasibility? My sister works in an area of the govt where they get some form of audit every year, and once it was for porn. She (geophysicist) had several false positives (pink rocks).

Answer 8

There is significant, recent research into detection of pornography using conventional classification methods. Examples are available here, and here.

Answer 9

• Is this ethical?

Depends on the implementation and reasonable expectations of the employees. For example, if your software scans any machine connected to the network, then there's an additional requirement that infra needs to prevent unauthorized machines from plugging in. (Maybe that should be obvious, but it's frequently overlooked on networks I've seen.)

• Is it feasible? I've done a lot of image processing/indexing but this seems like a whole new world of complexity.

Is it feasible to drug test every employee? Maybe so, but I question its worth. I would randomize it. Let employees know their machines may be scanned for inappropriate content at any time.

• Any references to successful techniques for discovering porn?

I'm not touching this one. I don't think I could keep my sense of humor in check. But watch out for The Scunthorpe Problem when searching text.

• Is it appropriate for me to archive the results when something is discovered?

This one concerns me the most, and I would ask a lawyer. I suspect if you find illegal content you may technically be legally obliged to disclose it. That's bad, particularly if the user was exposed by no real fault of his own. You(r client) will need real legal advice on how to handle this. Get HR and the lawyers involved.

Answer 10

From a purely technical standpoint: This sounds like an object category recognition problem. I've never done anything like that, but from what I've read, state of the art category recognition systems work like this:

• First you search for a large number of interest points (e.g. using a Harris Corner Detector, extremal points of LoG/DoG filters in scale space; some authors even suggest picking random points)
• Then you apply a feature transform to each point (something like SIFT, SURF, GLOH or many others)
• Combine all the features you found into a histogram (Bag-Of-Features)
• Use standard machine learning algorithms (like support vector machines) to learn the distinction between object categories using a large number of training images.

Answer 11

Everyone's computer operating system was probably installed from a disk image.

1. start with the disk image and get a list of files you probably don't need to scan.
2. get a list of all the other files on each PC.
3. pull the actual files from 10-20 random machines and use as a test bed
4. search for items in a profanity and questionable word (hotties, jugs, 'barely legal', joke, etc) dictionary
5. View the video - should anyone have any video at all?
6. View photos
7. Any video or image files that are questionable can be used to search the other machines

It will take one or two employees to get caught before any one will put anything on their work computer.

Charge an obscene amount of money for this service. I'll be Zappos would never do this to their employees.

Answer 12

Assuming you are a domain admin on the network.

1. C$ into each users desktop machine.
2. Copy porn files into personal private share.
3. Delete from original location.
4. Make popcorn.
5. Complete detailed analysis of all "evidence".

Answer 13

I just wanted to comment, but only have 1 rep, so I can't.

In the case of Gravatar, you could add a function to filter out from a list of clean sites in internet cache locations. I.E. Gravatar and other sites you don't want false positives from. You could also filter out things like the desktop wallpaper. If they are displaying porn on the desktop you'd think people would notice outside of your audit.

Answer 14

Such things never work reliably. You can use a blocklist to block domains either on name or on being included on some list (a common practice). But those lists are never complete, and blocking on name based on criteria can lead to many false positives.

You can block on words appearing in the text of sites, but again this can lead to false positives (and gets very slow as you need to parse every single bit of data that passes through your network in order to detect "naughty bits").

you can block pictures (and maybe sites containing them) that show more than a certain percentage of skintones. But again it leads to many false positives. A university medical department blocking a medical encyclopedia with images of limbs and torsos showing wounds and skin conditions is a well known example of that. And of course it'd be racist as it'd only block certain skintones. If you block colours matching Caucasian skin, there's always porn using black actors for example.

Best just trust your employees, and have policies in place for when that trust is broken.

Answer 15

I don't know, there has to be a middle answer, that isn't as invasive, but solves the real issue, LIABILITY.

Have them sign a waiver, that releases the company of any liability for illegal stuff found on work pc's, that is not work related.

Answer 16

1. tell the user a url is considered adult - bluecoat proxy does that.
2. License the thing google does in their image search http://code.google.com/apis/safebrowsing/ http://www.google.com/search?q=google+image+recognition+api
3. scan the computer for items not in a pre-agreed list.

Answer 17

Image and content analysis to determine the differences between a tasteful photograph of a person, a swimsuit photograph, a nude photograph, depictions of pornography... as far as I know is nowhere near sophisticated enough to do in software alone.

Fortunately crowdsourcing should be useful here, as @ammoQ suggested in a comment. However I don't believe members of 4chan or any other forum would appreciate the vast number of non-pornographic images, such as generic web graphics for buttons, frames, advertisements, etc. being posted.

My recommendation would be to look into existing crowdsourcing solutions, such as Amazon Mechanical Turk. (However the terms of service may explicitly prohibit the involvement of pornographic content, so be advised you might have to find another solution or roll your own.)

To make crowdsourcing feasible, your software should be prepared to do some or all of the following:

• Store information that links the content with the computer it came from
• Identify exact duplicates across the entire inventory and remove them (but origin information is retained)
• Downsample images to some dimension, perhaps 320x200, which is sufficient to identify the content of the image without retaining unnecessary detail and wasting storage space/bandwidth
• Create still images of video content at some regular interval and apply the same downsampling rule

Finally, the database of reduced images that represent the original image and video content is checked by users (or a designated team if you have the resources) according to your company's code of conduct. The program or interface might show a single image at a time, or a screen of thumbnails--whatever you deem best to obtain accurate information.

The identity of the computer from which images came should absolutely be secret and unknown to the persons evaluating the data. Additionally it should be randomized and each image probably checked more than once to remove bias.

The same technique could be used for text, but first the content could be scored by keyword rankings which remove the bulk of text from crowdsource review. Classifying a long document will of course be more time consuming than classifying an image.