X

Police blotter: Detecting computer-generated porn?

In this week's installment, the FBI claims that one of its analysts can spot computer-generated images just by looking at them.

Declan McCullagh Former Senior Writer
Declan McCullagh is the chief political correspondent for CNET. You can e-mail him or follow him on Twitter as declanm. Declan previously was a reporter for Time and the Washington bureau chief for Wired and wrote the Taking Liberties section and Other People's Money column for CBS News' Web site.
Declan McCullagh
5 min read
"Police blotter" is a weekly News.com report on the intersection of technology and the law.

What: FBI claims that one of its analysts can simply look at a photo and detect whether it's been altered in Photoshop or generated by a computer.

When: U.S. District Judge Nancy Gertner in Massachusetts ruled on August 11 and November 22, 2006.

Outcome: FBI's claim was rejected and its expert was not permitted to testify.

What happened, according to court documents:
In 2002, the U.S. Supreme Court overturned a federal law banning the possession of images of minors in lascivious poses that were either Photoshop-altered adults or completely computer-generated. Since then, to secure a conviction, prosecutors must prove that a defendant possessed images of real--not virtual--children.

This brings us to the case of Rudy Frabizio, whose employer discovered sexually explicit images on Frabizio's computer that appeared to involve minors. The FBI was contacted, and Frabizio was indicted on one count of possession of child pornography.

Photo hoaxes

Initially, the FBI chose as its expert witness Hany Farid, a Dartmouth College professor of computer science, who had written a program to determine whether an image was real. But then Frabizio's defense attorney discovered that the program had a 30 percent false-positive error rate: it frequently classified a real photograph as computer-generated. It also classified an image of a cartoon dragon called "Zembad" as real.

The U.S. Department of Homeland Security, which shares responsibility with the FBI for prosecuting child pornography, has paid for Farid's research on image detection. A DHS-funded technical report that Farid published after the Supreme Court's ruling claims his statistical technique "correctly correctly classified 67 percent of the photographic images."

After that revelation, the FBI quickly switched witnesses. Its new expert was Thomas Musheno of the FBI's Forensic Audio, Video, and Image Analysis Unit.

The FBI claimed that Musheno could simply look at each image--with no computer program required--and figure out which is legal and which is not. Musheno concluded that 6 of the 19 JPEG images definitely depict real children and 10 others "appear to be" real children. (Musheno holds a bachelor's and a master's degree in photography, not in any technical disciplines, and the FBI handbook (PDF) does not discuss how to detect computer-altered images.)

In response, Frabizio's defense counsel essentially argued that the line between real and virtual had disappeared. As evidence, the defense cited a computer-generated image of a woman in a fetal position, a "photo" of actress Jennifer Garner, and an image of a woman kneeling on a bed.

U.S. District Judge Nancy Gertner in Massachusetts wrote two opinions, the first on August 11 and the second on November 22. (A preliminary ruling is here).

Gertner seemed to take a dim view of the FBI's witness-switching and its claim that one of its analysts could simply look at a sexually explicit image and say confidently that it was of someone who's 17 years old and a minor--or 18 years old and an adult.

"I have serious doubts as to whether a person visually studying the images in this case can distinguish real pictures from manipulated or wholly virtual ones with the level of confidence required in a criminal prosecution," Gertner wrote. She cited computer science research that said even "experts cannot know whether a digital image is real or virtual."

Gertner ruled in August that the government was welcome to find an expert on computer-based graphical manipulation--but because Musheno was not one, he would not be allowed to testify. In November, she rejected the government's request to reconsider, which had argued that other federal circuits permitted such testimony from nonexperts.

Excerpts from Judge Gertner's August opinion:
I find that visual observation alone is inadequate to the task of evaluating the images in this case. If photographic experts as a general matter are inadequate to the task of identifying computer-generated images, then no level of experience in that field will suffice to qualify one as an expert. Indeed, allowing Musheno to testify would be like allowing a dentist to identify the causes of glaucoma. If the government offered an expert who eliminated the possibility of such imagery in this case, then Musheno's testimony might be admissible.

In a world of rapidly changing technology, where the availability and use of Photoshop and other, similar programs is widespread, substantial evidence suggests it may be possible to digitally create or manipulate photographs in a manner the naked eye cannot detect. The government has not shown otherwise. Under these circumstances, it is unreasonable to expect a lay jury to differentiate the real from the computer-generated. The government must therefore present an expert or other extrinsic evidence to prove that the images in question depict real children.

Whether the images in this case are real or virtual cannot be determined based on mere observation, however, even by a photographic expert. More specialized, computer-based knowledge is required to exclude the possibility that the pictures are wholly virtual. Furthermore, even if visual observation were sufficient, I would find that Musheno's qualifications and expertise do not justify the conclusions he proposes to make. I would allow him to testify to his observations, but not to his ultimate conclusions.

Excerpts from Judge Gertner's November opinion:
The government suggests that the Supreme Court has already resolved the issue in the case at bar. It did not. In Ashcroft v. Free Speech Coalition, the Court was addressing a hypothetical question and one from 2002 to boot. The Supreme Court did no more than assume that if the government's position were true, that virtual images are indistinguishable from real ones, that would be irrelevant to a constitutional analysis. The fact that it would be difficult to tell the real images--which are not protected--from the virtual images--which are protected--is not a basis for suppressing lawful speech, i.e. the virtual images.

The argument, in essence, is that protected speech may be banned as a means to ban unprotected speech. This analysis turns the First Amendment upside down. The Government may not suppress lawful speech as the means to suppress unlawful speech. Protected speech does not become unprotected merely because it resembles the latter. The Constitution requires the reverse. The possible harm to society in permitting some unprotected speech to go unpunished is outweighed by the possibility that protected speech of others may be muted.