Dmitry Shushlebin was sentenced on 16 September 2026 to six years in federal prison and $458,839.69 in restitution for a scheme that ran more than 100 fraudulent voter registration applications through the Pinellas County Supervisor of Elections. Coverage has filed it under election integrity. The Justice Department's own announcement calls it a synthetic identity theft fraud scheme, and that framing is the more useful one: the registrations were not the payload, they were the raw material for identities that went on to obtain credit cards and loans hundreds of times.
Which raises a question every identity verification team should be able to answer. When a government registration record turns up in an onboarding check as outside corroboration, what has anyone actually verified? Traced against the federal statute that creates it, the answer is narrower than the record looks: either an identity check performed by a different agency on a date the record does not carry, or, on one of the three available paths, no check at all.
- Shushlebin received six years, three years of supervised release and $458,839.69 in restitution after pleading guilty to four counts including wire fraud and aggravated identity theft.
- More than 100 fraudulent registrations were filed with Pinellas County in February and March 2023, and the Justice Department places them upstream of the synthetic identities, not downstream.
- Under 52 U.S.C. § 21083(a)(5) a registration takes one of three paths, and on none of them does anyone verify a living person at the time of registration.
- Where an applicant has neither a driver's licence nor a Social Security number, the State assigns a number instead, and each State decides for itself what counts as sufficient.
- The Federal Reserve made the same point about credit files in 2020: their existence "can seem to imply 'proof of life' and validation of the identity, but it does not."
- Triangulating across sources fails when a source is derived from the others, or authored by the applicant. Here it was both.
- Of the Federal Reserve's six synthetic-identity characteristics, one fires cleanly on this scheme, and it measures shared infrastructure rather than identity.
Dmitry Shushlebin, a 45-year-old Russian citizen living in Miami Beach, was sentenced on 16 September 2026 to six years in federal prison, three years of supervised release and $458,839.69 in restitution. The charge sheet that got him there ran through voter registration, but the Justice Department's own announcement does not call it an election case. It calls it a synthetic identity theft fraud scheme, and the distinction is the whole story: the more than 100 fraudulent registrations filed with the Pinellas County Supervisor of Elections were not what the fraud was for. They were what it was built from.
That inversion is worth sitting with, because it points at a weakness most identity verification stacks still carry. A registration record looks like outside corroboration of a person. Traced against the statute that creates it, it is nothing of the kind.
What the Florida court actually found
Shushlebin pleaded guilty on 25 July 2025 to four counts: conspiracy to give false information in registering to vote, wire fraud, making false statements, and aggravated identity theft. The Justice Department says he created sham companies to run a multi-year financial fraud scheme, used those companies to commit identity theft, and used the resulting fake identities to obtain fraudulent credit cards and loans across hundreds of transactions. He hired Sanjar Jamilov, an Uzbekistani national, and others to file the registration applications with Pinellas County in February and March 2023.
The investigation ran through the Federal Bureau of Investigation, the Florida Department of Law Enforcement and the U.S. Postal Inspection Service, and was prosecuted by the Criminal Division's Public Integrity Section alongside the Middle District of Florida. US Attorney Gregory W. Kehoe framed the outcome as a commitment to investigating and prosecuting threats to election integrity, and Assistant Attorney General A. Tysen Duva said that those who engage in fraud and deception in election processes will be investigated and prosecuted. Both statements are about elections. Neither is wrong. Both leave the financial mechanism unexamined.
Table 1: The documented sequence. Every entry is taken from a source we opened; nothing here is reconstructed from a headline.
Why sham companies came first
Order matters here. The companies came before the registrations, and the registrations came before the borrowing. That sequence is not unusual, and it is not improvised. The Federal Reserve described it in its July 2020 white paper Mitigating Synthetic Identity Fraud in the U.S. Payment System, which notes on page 3 that sophisticated crime rings leverage multiple tactics at scale to cultivate synthetic identities, including using fake addresses, creating sham businesses and forming relationships with collusive merchants to cash in.
A synthetic identity is not a forged document. It is a reputation, assembled slowly out of records that each look ordinary. The Federal Reserve's 2019 definition describes the combination of real information, such as a legitimate Social Security number, with fictional details, then the patient construction of creditworthiness before a bust-out. Shushlebin's sham companies were the workshop. The Pinellas filings were one of the components coming off the bench.
What a voter record proves about a person
FinCEN put the operative definition in a footnote of its November 2024 alert on deepfake media, and it is the sharpest one available: a synthetic identity refers to the use of a combination of real and fake personally identifiable information to fabricate a person or entity to pass validation processes. The phrase that matters is the last one. A synthetic identity is defined by the checks it survives, which means the useful question about any record is not whether it exists but which validation it lets an identity pass, and on what authority.
So we traced it. Voter registration for federal elections in the United States is governed by the Help America Vote Act, now codified at 52 U.S.C. § 21083. Subsection (a)(5) sets out exactly what an applicant must supply and exactly what the State does with it. There are three routes through, and the trace below follows each one to the party whose identity check is being relied upon.
Help America Vote Act · 52 U.S.C. § 21083(a)(5)
What a voter registration record actually checks
Federal law gives a registration application three routes through identity verification. Each one is traced here to the party whose check is being borrowed, and to when that party last stood in front of a human being.
What the applicant supplies
What the State matches it against
Who last verified a live person
A current and valid driver's licence number
§ 21083(a)(5)(A)(i)(I)
The state motor vehicle authority's own database, under a matching agreement
§ 21083(a)(5)(B)(i)
The licensing clerk, at the counter, on the day the licence was issued
Borrowed · years earlierThe last four digits of a Social Security number
§ 21083(a)(5)(A)(i)(II)
The motor vehicle authority, which in turn queries the Commissioner of Social Security
§ 21083(a)(5)(B)(ii)
Whoever processed the SSN application, commonly in infancy
Borrowed · decades earlierNeither. The applicant affirms having no licence and no SSN
§ 21083(a)(5)(A)(ii)
Nothing. The State assigns the applicant a number that will serve to identify them
§ 21083(a)(5)(A)(ii)
No one, at any point, on this path
No originWhat the trace shows
On none of the three paths does anyone verify a living person at the moment of registration. Two of them re-present an identity check performed by a different agency, for a different purpose, at a date the record does not carry. The third performs no check at all and mints an identifier anyway. Whether the numbers supplied are “sufficient” is left to each State to decide under its own law, at § 21083(a)(5)(A)(iii).
Three paths, and what each one borrows
Under § 21083(a)(5)(A)(i), an application may not be accepted or processed unless it includes either the applicant's driver's licence number, where one has been issued, or the last four digits of the applicant's Social Security number. Those are paths A and B, and both of them are lookups rather than checks.
Path A resolves under § 21083(a)(5)(B)(i), which requires the chief State election official and the State motor vehicle authority to enter into an agreement to match information in the statewide voter registration database with information in the motor vehicle authority's database. What that confirms is that a licence number exists and agrees with a name. The moment a human being was actually looked at is the moment the licence was issued, at a counter, on a date the voter record does not carry.
Path B resolves one hop further out. Under § 21083(a)(5)(B)(ii), the motor vehicle authority enters into its own agreement with the Commissioner of Social Security. A last-four match tells you that a number, a name and a date of birth agree with the Social Security Administration's records. For most people the underlying enrolment happened in infancy. That is a very old check being re-presented as a current one.
Neither path is useless. Both are genuinely good at catching a number that was invented rather than stolen. What neither does is establish that the person filing this application today is the person those upstream records describe, and an onboarding system that treats the resulting record as a second opinion has quietly counted the same opinion twice. The same structural problem shows up in document checks, which we traced in how stolen ID scans bypass document verification: a check that interrogates an artefact cannot tell you who presented it, a problem the IDScan breach of 153 million licence scans made concrete at scale.
When the State simply assigns a number
Path C is the one worth reading twice. Subsection (a)(5)(A)(ii), headed "Special rule for applicants without driver's license or social security number", provides that if an applicant has not been issued a current and valid driver's licence or a Social Security number, the State shall assign the applicant a number which will serve to identify the applicant for voter registration purposes.
Florida implements this directly. Section 97.053(5)(a)(5) of the Florida Statutes says that where an applicant has not been issued a current and valid Florida driver licence, Florida identification card or Social Security number, the applicant shall affirm this fact in the manner prescribed on the uniform statewide registration application. The affirmation is the applicant's own. There is nothing on the other end of it to match against, because the premise of the path is that no such number exists.
There is a further loosening one clause down. Under § 21083(a)(5)(A)(iii), the State determines whether the information an individual provides is sufficient to meet the requirements of the subparagraph, in accordance with State law. There is no federal floor on sufficiency. Fifty jurisdictions each answer the question for themselves, and a record produced under any of their answers looks identical to a downstream consumer.
We do not know which of the three paths Shushlebin's applications took. No source we read says, and it does not change the argument, because all three fail the same test in different ways. Two borrow their assurance from elsewhere without carrying the date. The third has no assurance to borrow.
How the corroboration turned circular
Put the statutory trace next to the sequence the Justice Department describes and something uncomfortable falls out. The registrations were filed by the operation. The county recorded them. The records then existed as official entries that any outside party could point to. And the identities those records described went on to obtain credit hundreds of times.
United States v. Shushlebin · Middle District of Florida
The record that vouched for the man who wrote it
Read as election fraud, the Pinellas registrations look like the crime. Read as identity fraud, they look like the raw material. The chain below is the order the Justice Department describes, with the onboarding check's field of view marked.
Step 2
A county office files it
The Supervisor of Elections receives applications in the ordinary way, against the checks traced above.
Step 3
It becomes a government record
A name and address now sit in an official register that any outside party can point to.
Step 4
A lender reads it as evidence
Credit cards and loans are obtained on the identities, hundreds of times over.
Credit grantedEverything a records-based check can see
The register and its contents. Not who filed them, not why, not whether one party filed a hundred.
Steps 1 and 4 are the same party. DDG's reading, not a finding of the court: a register entry offered as outside corroboration of a person is, on this chain, that person's own earlier assertion carrying a county seal.
Why this defeats triangulation
Checking one identity against several sources only helps when the sources are independent of the applicant. Here the fourth source was authored by the person being checked, and nothing in the record it produced carries that fact forward. The Federal Reserve made the same point about credit files in 2020: their existence “can seem to imply ‘proof of life’ and validation of the identity, but it does not.”
The corroboration and the thing being corroborated have the same author. A verifier looking at a public register sees a name, an address and a date of entry. It does not see who filed the application, why, or whether one operation filed a hundred of them in six weeks. The register is not lying. It is simply not a witness, and it has never claimed to be.
The Fed said this about credit files
The clearest articulation of this error in the American literature is already on the record, one layer over. On page 17 of the same 2020 white paper, the Federal Reserve observes that the Fair Credit Reporting Act requires a credit reporting agency to create a credit file when a credit inquiry is submitted, whether or not the customer's identity has been validated. It then draws the conclusion in one sentence: the existence of a credit file "can seem to imply 'proof of life' and validation of the identity, but it does not."
Read that as a general principle rather than a fact about credit bureaus and it covers the case at hand exactly. A record created on request is evidence that a request was made. Institutions repeatedly read it as evidence that a person exists, because the record is held by a serious organisation and looks official, and neither of those properties has anything to do with whether anyone checked.
The same paper is blunt about when the error surfaces. Page 4 states that the most likely point of detecting a synthetic identity is when a fraudster applies for credit. That is the moment the money is already in motion.
Britain built this into credit files
In one large market the inference has been formalised rather than merely assumed. Experian's UK consumer guidance on the electoral roll tells readers that when you register to vote, your electoral details are recorded on your report, that this data helps lenders confirm your name and address, and that your score will increase as a result. It adds that if lenders cannot confirm your details via the electoral roll, they may ask for other forms of identity and proof of address, and that electoral details are used to identify people by organisations beyond lenders. Equifax UK puts it more strongly still, describing registration as one of the most reliable ways of verifying your identity, and saying the register allows any interested party to confirm that you are who you say you are and that the details you have provided are accurate.
That is a clear statement that a registration record functions as an identity and address input to a credit file. It is also, on the trace above, a statement that an institution is treating a record with borrowed or absent assurance as a source of assurance in its own right. The Pinellas filings were American and no source we read says a US credit bureau ingested them, so nothing here asserts that Shushlebin's registrations reached a credit file by that route. The UK arrangement matters because it shows the inference written down, in a jurisdiction where a manufactured registration would convert into credit-file strength by design rather than by accident.
Table 2: The gap between what a signal proves and what it is scored as proving. Only the last row binds anything to the moment of the check, and it carries its own conditional.
Where triangulation quietly breaks
The standard answer to a weak signal is more signals. The Federal Reserve's paper cites the Better Identity Coalition's 2018 blueprint, which recommended triangulating data from multiple sources to validate individuals, for instance examining data from both the Social Security Administration and state motor vehicle databases, in order to avoid relying on personally identifiable information from a single source. That is sound advice and it is the architecture most identity stacks are built on.
Triangulation carries two assumptions, though, and they are rarely tested separately. The first is that the sources are independent of one another. A voter registration record is not independent of the SSA or the DMV; under § 21083(a)(5)(B) it is derived from them. Counting it as a third leg is counting the same evidence twice under a different name.
The second assumption is that the sources are independent of the applicant. That one failed completely here. A source authored by the person under examination is not corroboration at any volume, and adding two more sources of the same kind makes the score go up while the evidence stays where it was. This is the same shape of failure we described in how deepfakes bypass KYC, where each individual check performs as specified and the combination still lets an attacker through.
Why the detection window closes late
Six characteristics of synthetic identities appear in the Federal Reserve's second white paper: a credit file depth inconsistent with customer age or other profile information, multiple identities sharing one SSN, multiple applications from the same phone number, mailing address or IP address, use of secured credit lines or piggybacking to build credit, an SSN issued after 2011, and multiple authorised users on one account. The paper is careful to warn that focusing on any one characteristic alone produces false positives and disadvantages legitimate customers with short credit histories, such as recent immigrants.
We ran those six against the public record of this case. The result is not flattering to the list.
Table 3: The Federal Reserve's six published synthetic-identity characteristics, tested against what the public record of this case actually contains. One fires cleanly, and it is not a test of identity.
One of the six fires cleanly, and it is the one about shared phone numbers, addresses and IP addresses. That is a correlation test across applications, not a test of any individual identity, and it is the signal that a hundred filings from a single operation inevitably produces. It is also, notably, a signal that was available at the county office long before it was available at any bank. The detectable property of this scheme was its volume, not its people.
Everything else on the list is either downstream of the credit application or unestablished in the public record. For a detection programme, that is the finding: the characteristics most often cited are largely post-hoc, and the one with early reach is an infrastructure signal that no single onboarding decision can see. Prosecutions reach the same point late for the same reason, as the Taipei voice-conversion indictment and the PSNI investment-ad case both show.
What FinCEN still recommends
FinCEN's alert FIN-2024-Alert004, issued 13 November 2024, is the most recent US federal guidance addressing how fabricated identities get past onboarding. Its list of best practices is short. Multifactor authentication, including phishing-resistant multifactor authentication. And live verification checks in which a customer is prompted to confirm their identity through audio or video.
That second item is the only control in the whole chain that asks a person to be somewhere at a particular moment. It is the one thing a manufactured paper trail cannot pre-seed, because the record does not answer the phone. FinCEN then adds the caveat that defines the current threat: illicit actors may be able to respond to live verification prompts or access tools that generate synthetic audio and video responses on their behalf, and their responses may reveal inconsistencies in the deepfake identity.
So the control that survives records-based manipulation is precisely the control synthetic media is built to attack, which is why the two problems are converging. The attacker who can no longer buy his way past a data check will try to generate his way past a liveness check instead, and the mechanics of that shift are what we set out in how deepfakes bypass liveness checks and presentation attacks versus injection attacks. Detecting a generated response is a different discipline from validating a record, and an institution that has only ever done the second will find the first is not an upgrade of it. We cover the detection side of that in how AI impersonation attacks are detected.
What an onboarding team should change
Five practical changes follow from the trace, and none of them requires buying anything before Monday.
Score a record by the origin of its assurance, not by its existence. For every third-party source in the stack, ask which party last verified a living human and on what date. Where the answer is "a different agency, at a date this record does not carry", the source is a pointer, not a witness, and it should not be weighted as an independent confirmation.
Separate existence evidence from person evidence. Government-adjacent records are genuinely good at telling you that a name and an address have been associated before. They are not evidence that the applicant in front of you is that person. Most scoring models collapse the two, and the collapse is where a manufactured record earns its value. The equivalent failure in hiring checks is set out in how forged photo IDs bypass hiring verification.
Test for source independence, not source count. Two questions per source: is it derived from another source you are already counting, and could the applicant have caused it to exist? A stack of five sources where three descend from one upstream database and one was filed by the applicant is not a stack of five.
Move the binding check earlier. The only control that ties a claim to a person at the moment of the claim is a live capture whose path can be trusted. That means proving the image reached the system from a real sensor rather than a file or a virtual camera, which is a different property from whether the face in it looks genuine. We have written about where that boundary sits in how anti-spoofing works and where it stops working and liveness detection versus deepfake detection.
Correlate across applications, not just within them. This scheme's detectable property was that one operation produced more than a hundred records through shared infrastructure. Per-application scoring is structurally blind to that. Cross-application correlation on device, network and address is the control that would have seen it, and it is the one the Federal Reserve's own characteristic list quietly depends on. High-volume markets face the same arithmetic, which we covered in how deepfakes break eKYC in high-volume markets.
DuckDuckGoose works on the last two of these: DeepDetector is built around capture-path and generated-media questions at the moment of onboarding rather than around the documentary record an applicant brings with them.
What this article could not verify
Several things in this story are not established, and the argument is constructed so that none of them is load-bearing.
The Justice Department does not explain the mechanism by which the voter registrations helped create the synthetic identities. Its announcement says the identity theft allowed Shushlebin to create synthetic identities that appeared real after the registrations were submitted; ABC7 WWSB renders the same point as the registrations having been used to help create the identities. Both place the registration upstream of the identity becoming usable. Neither describes what consumed the record. Our reading of that chain is an inference from the sequence, and it is labelled as such in the second visual above.
No source we read names the lenders, describes the credit-building method, states whether real Social Security numbers were reused across identities, or says which of the three statutory registration paths the applications used. Nothing here should be read as a claim that any particular institution treated a voter record as corroboration in this case. The claim is about what the statute makes possible and what the record therefore cannot prove, which is checkable independently of what any one lender did.
DuckDuckGoose has not analysed any media or record connected to this case, and nothing in this article reports a detection result.
Methodology and what we checked
Incident facts come from the Justice Department's announcement of the sentencing, corroborated against ABC7 WWSB and The Daily Signal, both read directly. The Tampa Bay Times report was opened but is paywalled beyond its opening, so it is cited only for the sentencing date, the courthouse and the count of registrations, which are visible above the paywall. The Washington Times and WTSP both carried the story and both returned HTTP 403 to our fetcher, so neither is cited and no detail from either appears here.
The statutory trace is built from primary text. 52 U.S.C. § 21083 was read in the 2023 edition of the United States Code published by the Government Publishing Office, and cross-checked against Cornell's Legal Information Institute. Florida Statutes § 97.053 was read in the 2025 edition published by the Florida Senate. Subsection letters are quoted as they appear in those current editions; statute PDFs retrieved by search carry the numbering of the year they were printed, which is a trap this pipeline has walked into before.
The Federal Reserve material is quoted from the July 2020 white paper with page numbers, and from the Board's 2019 press release. The FinCEN definition and best practices are from alert FIN-2024-Alert004 dated 13 November 2024, read directly. The Social Security Administration's own documentation on its consent-based verification service would have strengthened one paragraph, but ssa.gov returned HTTP 403 to every URL we tried, so no claim in this article rests on it.
Frequently Asked Questions
Was this an election fraud case or an identity fraud case?
Both, legally. Shushlebin pleaded guilty to conspiracy to give false information in registering to vote alongside wire fraud, false statements and aggravated identity theft. The Justice Department's own headline describes it as a synthetic identity theft fraud scheme. The financial counts are where the $458,839.69 restitution comes from.
Does a voter registration record verify someone's identity?
No. Under 52 U.S.C. § 21083(a)(5), a registration either matches a driver's licence number against a state motor vehicle database, matches the last four digits of a Social Security number through that authority to the Social Security Administration, or, where the applicant has neither, results in the State assigning a number. None of the three involves anyone verifying a living person at the time of registration.
What is a synthetic identity?
FinCEN defines it as a combination of real and fake personally identifiable information used to fabricate a person or entity in order to pass validation processes. The Federal Reserve describes the same thing as combining real information, such as a legitimate Social Security number, with fictional details, then building creditworthiness over time before a bust-out.
Why does adding more data sources not fix this?
Because triangulation assumes the sources are independent both of each other and of the applicant. A voter record is derived from the same motor vehicle and Social Security data other checks already use, so it is not a genuinely separate source, and in this scheme it was filed by the person it was later used to describe.
What actually catches a scheme like this?
On the public record, volume did. More than a hundred applications from one operation through shared infrastructure is a cross-application correlation signal. Going forward, FinCEN's recommended control is a live audio or video verification check, which is the one thing a manufactured paper trail cannot prepare in advance, and which is why deepfake detection and identity verification are now the same problem.
About the author
