GENIUS, billionaire, camera-shy: there are a few things we can say about Satoshi Nakamoto, the founder of bitcoin – but we still don’t know who he, she or it actually is. Nakamoto has shaken up the financial world, but has not been heard from since 2011, and no one has definitively identified the person behind the name – though many have tried. Now come fresh claims that the world’s most elusive billionaire has been unmasked. All thanks to their writing style.
In 2014, a group of students at Aston University, UK, led by forensic linguist Jack Grieve, , published in 2008. They found clues to the writer’s identity in the frequency of innocuous words like “still” and “only”, and in punctuation patterns such as the use of commas before “and” and “but”. These matched the writing style of cryptocurrency polymath Nick Szabo.
Advertisement
And earlier this year, American entrepreneur and political pundit Alexander Muse claimed , although whether the analysis also pointed to Szabo is not public knowledge.
These efforts to chase down Nakamoto raise interesting questions about how we reveal our identities every time we write something. How much can be discerned about an author from the way they write? As digital communication proliferates, what are the clues hidden in our tweets, emails and messages that might give us away? And with the rise of software that can analyse masses of data and look for patterns, is there any way to hide?
Every time we speak or write, we shed huge amounts of information about who we are, what we do and where we come from. Detectives have used the written word to solve crime for centuries, but in the last few decades, computers have taken on some of the heavy lifting, analysing patterns in the swathes of digital information we now churn out.
This stylometric analysis is standard fare in undergraduate computer science classes – and is at the heart of plagiarism policing software used every day by universities and publishers, as well as by experts trying to identify criminals from their written traces online.
Most often, the aim is either to figure out precisely who wrote some text or to identify traits of the unknown author: their age, gender, education level or native language. The analysis usually begins with a line-up of possible authors and samples of their writing, from which experts or software extract highly salient features. The mystery text is compared with these, and the match, whatever it might be, is stated as a probability, not a definite yes or no. Those results are then considered along with other pieces of evidence that make up the case, such as alibis. “If all of these point in the same direction, then you can make a strong determination of who the author is,” says , a computational linguist at the Illinois Institute of Technology.
Devil strip detail
What counts as a salient linguistic feature can vary from case to case. A famous example involved a ransom note demanding that the money be left on the “devil strip”. Asked by police to help out, linguist at Georgetown University in Washington DC happened to know that “devil strip” was an extremely rare name for the grassy area between the pavement and the street – so rare, in fact, that only people in Akron, Ohio, use it. When Shuy asked the police if they had a suspect from Akron, their jaws dropped. They did, and the subject eventually confessed.
If you don’t want your words to unmask you, simply avoiding regionalisms or other kinds of special vocabulary won’t do. This is because the giveaway clues are usually commonly overlooked text features that we don’t consciously control: which words are capitalised, how many spaces we use after punctuation marks, whether paragraphs are indented. “You almost never get a devil strip,” Argamon says. “That was a really lucky break.”
Even more telling than text features are the relative frequencies of the so-called function words that glue together sentences. “These don’t carry meaning on their own and serve grammatical functions, words like prepositions, conjunctions and personal pronouns,” says Argamon. One reason they are so rewarding to analyse, at least in English, is their sheer number: altogether, they make up more than half the words we write (see “Literary secrets laid bare”).
These seemingly meaningless bits of language could point to someone’s personality type, health outcomes and even future acts such as suicide, according to work by James Pennebaker at the University of Texas at Austin.
Others are exploiting telltale linguistic quirks in an effort to infiltrate criminal networks online. , also at Aston University, trains undercover police officers to assume the online identities of arrested paedophiles, in order to catch others or pose as potential victims. Grant creates portraits of their writing style to train officers to monitor their own mimicry. “These are low-trust environments and people are very alert to the wrong language,” he says. “If you get the vocabulary choices or communication behaviours wrong, the interaction can get uncomfortable.”
These salient patterns, whether in word choices, sentence structures or unconscious frequencies of function words, point to the remarkable flexibility of language. Linguists used to believe that we all learn a single uniform grammar of a language, then begin to deviate from it to express our personalities. More prevalent now is the idea that we each possess a mental model of our own language, one that differs in slight but important ways from those of others, due to the social and emotional context in which we learned it.
This leads to individual ways of writing, according to , a computer scientist at the University of Arizona, who first articulated the notion of a “writeprint”. The linguistic equivalent of a fingerprint, a writeprint is made up of subtle differences in our writing style, such as vocabulary, sentence length, layout of paragraphs and so on, Chen says.
Disappearing in style
So how can you evade those trying to take your linguistic fingerprint? In one sense, it’s easy, Argamon says. Imagine there are 100 authors, who have each submitted a text, and that you don’t want anyone to be able to tell which one you wrote. “All you have to do is make your text look like one of the other 99 authors,” he says.
This works well in some cases. In one study, , now at the International Computer Science Institute in Berkeley, California, and her colleagues at Drexel University in Pennsylvania asked people to write about their morning as if they were the US writer Cormac McCarthy. A stylometric program that had already been trained on McCarthy’s writing was fooled: it thought they were the real thing.
Unfortunately, people generally prove to be enduring amateurs at identifying the right changes to make. In the “devil strip” case, the writer of the ransom note deliberately misspelled some words (like “kop” for “cop” and “kan” for “can”) to disguise himself as a less educated person. Unfortunately for his linguistic subterfuge, he spelled some difficult words correctly, an inconsistency that revealed the ploy.
“If an individual tries to obfuscate through writing by making it look like someone different from themselves, they’ll usually inadvertently create even more features that can be traced back,” says Argamon. Tim Grant recalls being involved, along with his colleague Jack Grieve, in the case of Jamie Starbuck in 2013. Starbuck spent nearly three years travelling the world sending emails that purported to come from his wife, Debbie – it turned out that he had murdered her 31 months earlier, a week after their marriage – but only began imitating her writing style when her relatives became suspicious. “She was a heavy user of the semicolon and would use it in idiosyncratic ways,” Grant says. “He suddenly started using many more semicolons, but he couldn’t bring himself to use them in the right way.” Starbuck was eventually arrested when he arrived back in the UK and sentenced to life in prison.
Could computers themselves be used to modify our writing and so outwit stylometric analysis? All you would do is upload your writing and a program then tells you what to change. The idea is now a burgeoning field called adversarial stylometry, and Afroz is one of the researchers trying to move things forward. Her involvement meant she was constantly being asked to unmask Nakamoto, she says, and eventually she put a note on her website saying she refused to do so. “My goal for working on stylometry is to make people aware of its harm, investigate the brittleness of machine learning and make tools to IMPROVE anonymity,” she wrote.
Preserving anonymity can be a legitimate professional interest, as when academics want their anonymous peer reviews of grants or papers to stay that way. Anonymity can also be a matter of life or death for whistleblowers, activists and even programmers (see “Code within code”). So the appeal of software that would anonymise text is obvious. But does it work?
So far, the only publicly available anonymising tool is , developed at Drexel University’s Privacy, Security and Automation Lab. Released in 2012, it aims to reduce the accuracy of stylometric analysis to no better than random guessing. To achieve this, Anonymouth uses a stylistic analyser called JStylo, which builds a profile of each particular writer from a given sample of texts by assessing features like sentence length, word choice and the frequency of certain letters. Then Anonymouth suggests how the writer can alter the text to make it look less like this profile – switching tenses from past to present, for example, or using certain pronouns less frequently.
JStylo’s creators claim that using 6500-word samples, it can match each text to its author from a set of candidates it has already been trained on, with 80 to 85 per cent accuracy. And earlier this year, an artificial intelligence program called Emma Identity was announced, claiming to need only 8000 words to build an author’s profile that was 85 per cent accurate at matching anonymous samples.
“People prove to be enduring amateurs at covering their tracks”
These success rates are far from perfect, but they are significantly better than chance. That’s possible because the analyses are run in lab-like situations with specific instructions about what kinds of stylometric characteristics to look for. In the real world, however, the writing samples available to train an analyser like JStylo or Emma may be too short, or they might be hastily typed emails when the anonymous text is a carefully written letter or scientific paper.
In future, we could see “adversarial authorship” – an accelerating arms race pitting technologies that identify writers against those that anonymise them. A tool called AuthorWeb, being designed at the North Carolina Agricultural and Technical State University, could enable writers to evade stylometric analysis. It gives them stylistic targets to hit as they write, using a visual dashboard that provides real-time feedback on how closely their writing matches certain features. This should help writers hide their style more easily and consistently over longer periods of time.
In the meantime, the most promising way to evade stylometric analysis is to write collaboratively, Argamon says: one person writes the text, someone else edits it. Rather than depending on a machine to alter your style or doing it yourself, this is a case of the linguistic fingerprints of two or more people cancelling each other out. And that may be exactly the tactic that has kept Satoshi Nakamoto hidden for so long: some suggest that a group, not an individual, lies behind bitcoin. With their linguistic fingerprints smudged together, they may safely continue to lurk, watching the chase go on.
Literary secrets laid bare
Revealing an author by analysing seemingly meaningless bits of language such as so-called function words has history. The technique came to prominence in 2013 when Patrick Juola at Duquesne University in Pennsylvania used similar methods to unmask J.K. Rowling as the author of the novel The Cuckoo’s Calling.
When the author is dead, it becomes harder to convince everyone, however. Take the 19th-century poem Twas the Night Before Christmas, historically attributed to Clement Clark Moore. In 2016, New Zealand literary scholar Macdonald Jackson published an exhaustive analysis that used words like “that” and “the” and pairs of phonemes to settle on Henry Livingston as the author.
This didn’t sit well with everyone, however. Scott Norsworthy, a specialist in the works of Herman Melville, derided Jackson’s (and his computer’s) use of “meaningless bits, inconsequential and perhaps random in their distribution”.
Code within code
It might look as functional as can be, but computer source code can give much away about the person – or group of people – who wrote it. Coders possess a “code print”, just as writers have a “writeprint”, because there are many ways to create a program.
“Depending on people’s comfort level and coding skill level, they choose different ways to do things,” says in Berkeley, California.
The code print can involve seemingly minor choices, such as indenting lines using the space bar rather than the tab key, each of which leaves a distinct digital trail. Even the most basic instructions to computers contain variations depending on who wrote them.
In 2015, a team of computer scientists at Drexel University in Pennsylvania used software to analyse the coding style of 1600 participants in the annual Google Code Jam. By looking at keywords and other aspects including the syntax of the code, the software successfully matched code to its author with nearly 93 per cent accuracy.
Analysing contributions over time, the team also found that coders’ styles hold steady over a span of several years. The stability of this code print could be important if the only available samples of a known author’s code date from some time before.
Why would programmers want to remain anonymous, anyway? We tend to think of the insidious case of the malware writer, but there are legitimate reasons for wanting to hide your tracks. For example, someone may not want to be known as a contributor to open source software if that software is illegal where they live. And activists who produce tools to circumvent censorship may not want to leave traces that would allow governments to identify them.
This article appeared in print under the headline “Write yourself invisible”
