Snap Shot
- The question: How do most child sexual abuse material (CSAM) investigations begin, and what can hash matches, CyberTipline reports, and peer-to-peer logs establish when the government relies on them?
- What the research shows: Independent testing found PhotoDNA robust to compression and resizing but much less robust to cropping; in one laboratory reimplementation no collisions between unrelated images were observed at the thresholds the author considered reasonable; a neural perceptual hash could be forced into false matches with changes that were often hard to see. Law enforcement software can log U.S. computers offering known files.
- What it does not show: A hash match shows that a file resembles one someone previously classified, not who placed it on an account, whether the user viewed it, or intent. An IP address identifies a connection, not a person. No independent field error rate for PhotoDNA was located.
- Where courts disagree: The Fifth and Sixth Circuits allow police to open a hash-matched file that no one at the provider viewed without a warrant; the Ninth, Second, and Fourth require one. Good faith, inevitable discovery, and attenuation prevented suppression in most cases summarized, although the Ninth Circuit reversed a denial of suppression in one Arizona case; in another it held that hash matches can supply probable cause without anyone viewing the suspect’s copies.
- Issues for each side: Prosecutors are best served by obtaining a warrant before opening unviewed files and proving hash reliability on the record; defense counsel should obtain the full CyberTip and hashing evidence, recognizing that a violation finding seldom ends in suppression.
Most federal, state, and military CSAM prosecutions do not begin with a victim’s complaint or an officer’s observation. They begin when an online service detects a suspected file, usually by comparing it automatically with a database of previously identified images, and reports it to the CyberTipline operated by the National Center for Missing & Exploited Children (NCMEC). Investigators then use the internet protocol (IP) address and account information in the report to identify a subscriber, a residence, and a device. Other cases begin when investigators observe a computer offering known files on a peer-to-peer network, or when an operation against an anonymous website identifies its users.
Each step can be challenged under the Fourth Amendment: whether opening a flagged file was a search, whether NCMEC or the provider acted for the government, whether subscriber records could be obtained without a warrant, and whether an affidavit established probable cause to search a particular home and device. Those challenges turn partly on technical questions about what a hash match, a CyberTip, or a peer-to-peer log can show. Statutes and courts use the term “child pornography,” and this paper uses that term when describing the law; elsewhere it uses CSAM.
This paper describes the detection systems and what the research says about their reliability, then summarizes the federal, military, and Arizona law. What files on a device show about the user’s knowledge is discussed in Digital Evidence and What It Can Show About the User; this paper concerns how the government reaches the device. The legal summaries are written by a non-attorney for orientation.
How Providers Detect and Report Suspected CSAM
Federal law requires electronic communication and remote computing service providers that obtain actual knowledge of apparent child pornography and related offenses to report them to the CyberTipline (18 U.S.C. § 2258A), but it does not require providers to monitor users or to affirmatively search, screen, or scan for such material (§ 2258A(f)). Many large providers scan voluntarily. The REPORT Act of 2024 added sex trafficking of minors and online enticement (18 U.S.C. §§ 1591 and 2422(b)) to the mandatory reporting categories, extended the period for which a report operates as a request to preserve the content from 90 days to one year, increased penalties for a provider’s failure to report, and limited liability for NCMEC-designated vendors and for people depicted as minors in such material, or their representatives, who report it.
Detection usually relies on hashing. A cryptographic hash, such as MD5 or SHA-1, matches only an identical file; re-encoding the image or altering a single pixel produces a different value. A perceptual hash, such as Microsoft’s PhotoDNA, is designed to produce similar values for images that look alike, so that resized or recompressed copies of a known image still match within a chosen distance threshold (Farid, 2021). A perceptual match is therefore a judgment that two images are close enough, not a determination that they are identical. Some providers also use machine-learning classifiers to flag material that is in no database; no previously classified image stands behind such a flag.
The statute leaves a report’s contents to the provider’s discretion: it may include identifying information such as email and IP addresses, when and how the content was uploaded or discovered (with date, time, and time zone), geographic information, the flagged images, and the complete communication containing them (§ 2258A(b)). Whether anyone at the provider viewed a file before reporting it is often indicated in the report, and that fact has become central to the law. NCMEC reviews reports, attempts to determine a geographic location, and makes them available to the appropriate law enforcement agency.
What a Hash Match Can and Cannot Establish
In the ordinary case, a match to a database image previously classified as CSAM is strong evidence that the flagged file depicts the same material. Independent research on the specific tools, however, is small; little detailed information about PhotoDNA’s performance had been published before Steinebach’s analyses (2023, 2024).
Steinebach (2024) reimplemented PhotoDNA and tested it on several image collections. Because the hash reduces each image to a 26-by-26-pixel representation, it was highly resilient to JPEG compression and resizing; at the threshold the author treated as a reasonable setting, about 98.7% of compressed or resized images still matched their originals. Cropping was different: removing one row and column of pixels had little effect, but when about nine rows and columns were removed, only about 16% of images still matched at that threshold (holding matches after that much cropping required a looser threshold of about 225 to 250), and the hash was not inherently robust to mirroring or rotation. He observed no collisions between unrelated images in his test sets at the thresholds he considered a reasonable compromise, although at looser thresholds similar photographs from the same series did collide (about 0.3% of pairs in one collection at a threshold of 225), and he noted that real databases are far larger, so thresholds may need to be stricter, and questioned whether the error rates claimed for the algorithm hold for cropped images. He concluded that PhotoDNA is a solid robust hash with room for improvement. The study used a reimplementation and laboratory images, not Microsoft’s production system, and did not estimate how often a deployed system produces a false report.
Other perceptual hashes can be deliberately manipulated. Struppek and colleagues (2022) attacked Apple’s NeuralHash, a neural-network hash proposed for scanning images on users’ devices, and were able to force hash collisions for about 91% (90.81%) of tested images and to change images’ hashes, using alterations that were often hard to see; they concluded that this kind of hashing was not robust enough for client-side scanning. Those findings concern a different algorithm and deliberate attacks, not chance error in PhotoDNA, but show that a perceptual match can in principle be defeated or engineered. Security researchers have also argued that on-device scanning creates security and privacy risks and offers law enforcement value that is at best problematic (Abelson et al., 2024), an expert position in a continuing policy debate rather than an empirical finding.
For a court, the practical points are these. A perceptual match shows that a file is visually similar, within a threshold the provider chose, to a file someone previously classified, and that classification is itself a human judgment; a cryptographic match shows that the file is identical to the database file. Neither shows who uploaded the file, whether the account holder viewed it, the age of the person depicted (see Estimating Age from Images), or intent. No published, independent estimate of PhotoDNA’s error rate in production was located, and the Second and Fourth Circuits have required the government to prove reliability with record evidence. Newly produced images, including AI-generated material, are by definition not in databases of previously identified files and can be flagged only by classifiers or human review, if at all.
CyberTipline Volume and Actionability
NCMEC received 36.2 million CyberTipline reports in 2023 and 20.5 million in 2024, attributing part of the decline to a “bundling” feature that lets a provider combine duplicative reports, such as many reports of the same viral file, and part to reduced reporting by some providers and to end-to-end encryption (National Center for Missing & Exploited Children [NCMEC], 2025). For 2025, NCMEC reported 21.3 million reports containing more than 61.8 million files, and more than three-quarters of reports involved users outside the United States (NCMEC, 2026a, 2026b). It also reported many reports with some generative-AI connection, most from a single provider and without actionable information (NCMEC, 2026b). The REPORT Act’s expansion of mandatory reporting contributed to a sharp rise in online enticement reports, so year-to-year changes reflect reporting rules as well as conduct (NCMEC, 2026a).
These figures show the scale of the problem and the triage burden; they say nothing about the accuracy of any particular report. A CyberTip is a submission, not a count of offenders, victims, or images, and it may be duplicative, concern a user abroad, or lack identifying information. No peer-reviewed study was located that measures what proportion of CyberTips lead to an investigation, a warrant, or an arrest.
Peer-to-Peer Investigations and IP Addresses
Investigators also monitor peer-to-peer networks with law enforcement software that identifies computers offering files whose hash values match known CSAM and logs the IP address, time, and files. Using a year of such observations on one network, Wolak, Liberatore, and Levine (2014) found that 244,920 U.S. computers shared 120,418 unique known files; more than 80% shared fewer than 10 such files or shared on fewer than 10 days, and fewer than 1% (915 computers) shared 100 or more. Bissias and colleagues (2016), measuring five networks from mid-2011 to 2014, estimated about 840,000 unique installations per month sharing such material worldwide; in their survey of investigators, 9.5% of people arrested for peer-to-peer trafficking were identified during the investigation as having committed contact offenses, ranging from 8% on one network to 21% on another.
Both sides draw on these studies. They show that law enforcement tools can document distribution by particular IP addresses with specificity, and the contact-offense figures may be offered in probable-cause or detention arguments. Their limits matter equally: they counted only files already known to investigators, so their estimates are lower bounds; the networks and period are now dated; and the contact-offense figures were reported by investigators, not adjudicated, and describe a group rather than any defendant. Detection route may also matter: among 336 Canadian men convicted of CSAM offenses, those identified through undercover policing were least likely to commit a new sexual offense over five years, and those identified during another investigation, usually of contact abuse, most likely, as reported in a book synthesis (Seto, 2025, Chapter 2) that does not replace the underlying peer-reviewed study. The relationship between online and contact offending is addressed in Online and Contact Offenders Compared.
An IP address identifies an internet connection at a moment, not a person. A residential connection may be shared by household members and guests, a wireless network may be open or compromised, and shared carrier addressing, virtual private networks, and anonymizing services can sever the link between an address and a subscriber. Few detected offenders appear to use such tools (16% and 22% in two samples, as reported in a book synthesis, Seto, 2025, Chapter 2), though those caught are likely less technically skilled than those who are not. Investigators address these possibilities through surveillance, interviews, and the forensic examination, and the strength of the link is a factual question in each case. Users of hidden services on the Tor network cannot be identified by IP address in the ordinary way, which led to the remote-access searches discussed below.
Limits of the Evidence
The hashing research consists of a few independent laboratory studies of reimplemented or proposed algorithms; it shows how the methods behave, not how often any provider’s system errs, and providers’ thresholds and classifiers are rarely documented publicly. The peer-to-peer studies measured particular networks in particular years, counting only known files, and their contact-offense estimates rest on investigators’ reports. CyberTipline volumes describe reporting activity, which is shaped by provider practice, encryption, and statutory change.
None of these sources can establish what a particular person did. A hash match, a CyberTip, or a log entry is a starting point for an investigation; its value in a given case depends on how it was generated, how it was corroborated, and what the search of the device found.
Questions to Ask About This Research
When a hash match is described as virtually certain, an attorney should ask how much evidence the reliability claim rests on: Steinebach (2024) tested a reimplementation of PhotoDNA on laboratory images and did not estimate how often a deployed system produces a false report. When the NeuralHash collisions (Struppek et al., 2022) are cited against a match, the question is whether the finding applies to this case, since it concerns a different algorithm and deliberate attacks. When the 9.5% contact-offense figure (Bissias et al., 2016) is offered in a probable-cause or detention argument, ask what exactly was counted: investigator-reported, not adjudicated, offenses describing a group. When CyberTipline volumes (NCMEC, 2025) are cited, ask what the count represents, since a report is a submission, not a count of offenders. These questions are explained in “Reading Research Claims: A Guide for Attorneys” on this site.
Search and Seizure in CSAM Investigations and the Law
Federal and Military Standards
Under the private-search doctrine of United States v. Jacobsen (1984), when a private party has already examined an item, the government may replicate that examination without a further search so long as it learns nothing more. The courts of appeals disagree about how the doctrine applies when a provider’s automated hashing flags a file that no person at the provider viewed and an officer then opens it. The Fifth Circuit in United States v. Reddick (2018) and the Sixth Circuit in United States v. Miller (2020) held that opening the file was not a separate search, reasoning that the hash match had already established its contents and that viewing revealed nothing significant beyond it. In Miller the defendant had not challenged the finding that hash matching was highly reliable, and the court left open how such a challenge should be decided; it also held that data NCMEC’s systems generated automatically was not a “witness” statement under the Confrontation Clause. The Ninth Circuit in United States v. Wilson (2021), the Second Circuit in United States v. Maher (2024), and the Fourth Circuit in United States v. Lowers (2026) held that a warrant was required. Wilson reversed the denial of suppression and vacated the conviction. Maher held that the provider’s terms of service did not defeat the user’s expectation of privacy and that the government must prove hash reliability, but affirmed under the good-faith exception because no appellate court had required a warrant when the file was opened in 2020. Lowers held that the district court erred in treating hash matching as exceedingly reliable through judicial notice rather than record evidence, and affirmed on attenuation grounds; one judge concurred in the judgment and would have followed Reddick and Miller. On October 5, 2026, the Supreme Court denied certiorari in Gasper v. Wisconsin, in which the Wisconsin Supreme Court had held that government viewing (first by a state Department of Justice analyst, then by a detective) of a video Snapchat had flagged with PhotoDNA did not exceed Snapchat’s private search, so the split remains.
The leading case from Arizona is United States v. Holmes (9th Cir. 2024), an appeal from the District of Arizona. A social-media provider reported two images to NCMEC without any employee viewing them, one matched by hash and one not, and an FBI agent viewed both without a warrant in early 2021, before Wilson. The government conceded that the viewing was a search requiring a warrant (slip op. at 12). The majority held that the good-faith exception for reliance on binding precedent did not apply, because the law at the time made the viewing at most plausibly permissible (slip op. at 12–18), and that inevitable discovery failed because the government’s parallel-investigation theory was speculative; it reversed the denial of the motion to suppress, which reached the evidence from the account and residence searches, the images found on the defendant’s phone, and his statements, and remanded (slip op. at 18–31). A dissenting judge would have held the viewing of the hash-matched image lawful; he agreed that viewing the unmatched image was unlawful but would have affirmed because the untainted information in the affidavit established probable cause (slip op. at 31–38, Collins, J., dissenting).
Whether hash matches alone supply probable cause is a different issue. In United States v. Johnsen (9th Cir. 2026), a peer-to-peer case from the District of Arizona, agents could not download the defendant’s files but matched the hash values of nine files he offered to known files and viewed library copies. The court held that the matches supported probable cause without visual confirmation of his own files, which would demand certainty rather than a fair probability, and that hash evidence need not be unassailable to be highly probative; descriptive file names and his prior convictions added support.
Whether NCMEC acts for the government is a separate question. In United States v. Ackerman (2016), the Tenth Circuit held that NCMEC is a governmental entity and, alternatively, that it acted as the government’s agent, and that its analyst’s opening of an email and viewing of three attachments the provider’s hash filter had not matched exceeded the provider’s private search. On remand the court affirmed under the good-faith exception (unpublished, 2020), and in United States v. Tolbert (2024) it assumed without deciding a violation by NCMEC but affirmed under inevitable discovery (slip op. at 14–15, 26–27). Providers, by contrast, have generally been treated as private actors even though they must report: the Eighth Circuit in United States v. Ringland (2020) held that the § 2258A reporting duty did not make the provider a government agent, and the Seventh Circuit in United States v. Bebris (2021) treated the provider’s scanning and review as private conduct, as did the Arizona Court of Appeals for Google in State v. Fristoe (2021), discussed below.
In 2015 the FBI operated a seized Tor-based CSAM website known as Playpen for about two weeks and, under a single Eastern District of Virginia warrant, deployed a “network investigative technique” (NIT) that caused visitors’ computers to send their true IP addresses and other identifying information to the government. The Eighth, Third, and Ninth Circuits held that the warrant violated the venue limits of Federal Rule of Criminal Procedure 41(b) (United States v. Horton, 2017; United States v. Werdene, 2018; United States v. Henderson, 2018), while the First, Fourth, and Tenth Circuits assumed a violation or did not decide the question (United States v. Levin, 2017; United States v. McLamb, 2018; United States v. Workman, 2017). Every court of appeals to decide the issue applied the good-faith exception and declined to suppress. District courts reached the same result; United States v. Anzalone (D. Mass. 2016), for example, found probable cause, called the Rule 41(b) question close, and held that good faith would apply in any event. Rule 41(b)(6), effective December 1, 2016, now authorizes remote-access warrants when technological means have concealed a device’s location.
Subscriber records stand on different ground. Carpenter v. United States (2018) held that acquiring seven or more days of historical cell-site location information is a search requiring a warrant, in a ruling the Court described as narrow, and Chatrie v. United States (2026) held that obtaining Google Location History through a geofence warrant was a search, extending Carpenter‘s reasoning to that location data and leaving particularity, probable cause, and good faith for remand. Neither decision addresses IP addresses or subscriber information.
In courts-martial, Military Rule of Evidence 311 excludes evidence from an unlawful search only if the accused moves in time, had a reasonable expectation of privacy or other basis to object, and exclusion would produce appreciable deterrence whose benefits outweigh its costs (Mil. R. Evid. 311(a)). It recognizes inevitable discovery, objective good-faith reliance on an authorization or warrant, and reasonable reliance on a statute or binding precedent (Mil. R. Evid. 311(c)), provides a hearing on a substantial preliminary showing of a knowing or reckless false statement or material omission, necessary to the finding of probable cause, in the information given the authorizing official (Mil. R. Evid. 311(d)(4)(B)), and places the burden on the prosecution once the defense moves (Mil. R. Evid. 311(d)(5)). Under Military Rule of Evidence 315, a commander, military judge, or magistrate may issue a search authorization, written or oral, on probable cause, and only a military judge may issue a warrant for wire or electronic communications under Rule for Courts-Martial 703A. (Rule text is from the 2024 Manual for Courts-Martial; Executive Order 14130, which postdates it, did not amend Military Rules of Evidence 311 or 315 or R.C.M. 703A.)
The Court of Appeals for the Armed Forces (C.A.A.F.) has addressed device searches in several cases, not all involving CSAM. In United States v. Nieto (2017), arising from cell-phone recordings in a latrine, an agent’s generalized profile of how service members move images from phones to laptops, without a particularized nexus to the accused’s laptop, did not support probable cause to seize it, and neither good faith nor inevitable discovery applied. United States v. Perkins (2019) assumed that an authorization lacked probable cause and applied good faith, which it held is met when agents reasonably believe the authorizing official had a substantial basis; United States v. Blackburn (2020), involving a hidden camera that recorded a child, assumed an insufficient nexus to the accused’s other devices and applied good faith. United States v. Richards (2017) upheld an authorization to search electronic media for communications with a minor without a date limit when read with the affidavit, and United States v. Ixcolgonzalez (2026) held that an authorization must limit the search to evidence of specific crimes or types of material but need not have a temporal limit, rejected a rule requiring as much specificity as the government’s knowledge allows, and, setting aside a lower court’s affirmance of suppression, held that the authorization met that standard and that it also incorporated the affidavit, whose categories supplied particularity. In United States v. Armour (2026), arising from a CyberTip, the military judge had suppressed an initial seizure as resting on a recklessly false affidavit that described an IP address NCMEC had listed only as a login address as one from which uploads were confirmed; the C.A.A.F. affirmed the service court’s reversal of the suppression of a later search on a corrected affidavit, holding that the independent-source inquiry ended once neither the decision to seek nor the decision to grant the new authorization was tainted. In United States v. Gurczynski (2017), examining a thumb drive for child pornography, an offense the warrant did not name, more than nine months after the warrant issued and five months after the accused’s conviction of the warrant offenses was unreasonable, and suppression was affirmed.
Arizona
Article 2, § 8 of the Arizona Constitution provides that no person shall be disturbed in his private affairs, or his home invaded, without authority of law. In State v. Mixton (2021), a 4–3 decision arising from a Tucson undercover investigation on a messaging application, federal agents obtained an IP address from the application and subscriber information from the internet service provider by federal administrative subpoena. The Arizona Supreme Court held that neither the Fourth Amendment nor the Arizona Constitution requires a warrant or court order for IP addresses and subscriber information, which may be obtained by a lawful federal administrative subpoena; the third-party doctrine still governs such non-content information after Carpenter, and the information is not a “private affair” under art. 2, § 8. Mixton concerns subscriber and IP records, not content or location data, and Chatrie does not by its terms disturb it.
A.R.S. § 13-3018 authorizes a prosecutor to subpoena a communication service provider for subscriber information, account identifiers, access logs, and routing records, but not content, in investigations of offenses involving a computer or network. A.R.S. § 13-3925 adds a statutory good-faith rule: a trial court may not suppress otherwise admissible evidence seized as a result of a “good faith mistake” or “technical violation,” terms the statute defines to include reasonable reliance on a warrant later invalidated because of a good-faith mistake and on controlling precedent later overruled. Under Arizona Rule of Criminal Procedure 16.2(b), the State must prove by a preponderance the lawfulness in all respects of its acquisition of evidence it will use at trial, but where the evidence was obtained under a warrant, or involves a search or seizure whose circumstances the defendant is entitled to discover under Rule 15, that burden arises only after the defendant alleges specific circumstances and establishes a prima facie case for suppression. In State v. Fristoe (2021), a published Court of Appeals decision arising from images Google found in a user’s photo account and reported to NCMEC, which forwarded the report to police, the court held that Google was not a government agent: § 2258A requires reporting but not searching, and Google had its own business reasons to monitor its platform. Assuming without deciding that NCMEC is a government actor, the court held that its review stayed within Google’s private search because it viewed the same images Google had viewed. Reviewing for fundamental error because the issue was raised for the first time on appeal, the court also held that the private-search doctrine applies under Arizona’s Private Affairs Clause, declining to follow the Washington Supreme Court’s rejection of the doctrine under that state’s constitution in State v. Eisfeldt (2008). A Google employee had viewed all but one of the images, and the court treated the claim concerning the unviewed image as moot because the defendant was acquitted on that count. Fristoe therefore did not decide how the doctrine applies to hash matches that no provider employee viewed, or whether a CyberTip alone supplies probable cause, and no published Arizona appellate decision on those questions was located; in federal court in Arizona, Wilson, Holmes, and Johnsen are binding. A later memorandum decision, State v. Ortega (2026), which does not create precedent and may be cited only for persuasive value under Ariz. R. Sup. Ct. 111(c), addressed a warrant based on a CyberTip. A messaging app had flagged files shared from the defendant’s account; police traced the IP address to a subscriber through a warrant to the provider and then obtained a warrant for electronic devices at his home. The Court of Appeals held that the CyberTip, the subscriber information, and the detective’s training-and-experience statement that people who possess CSAM store it on multiple devices supported probable cause to search electronic devices with digital memory, related accessories, and audio or visual files, not only the phone the app ran on, and that the warrant was not overbroad and did not need to be more particular (¶¶ 8–9); the court did not decide whether the good-faith exception applied. It also held that police could use a passcode the defendant’s domestic partner gave them after he refused to provide it, because by sharing it with her he had surrendered any expectation of privacy in it and, alternatively, the warrant authorized police to circumvent passcodes (¶¶ 15–17); that the trial court did not abuse its discretion in refusing to compel a test extraction on a prepopulated phone or further access to the Graykey extraction device, because the defendant had the phone and the raw extraction data and had not shown substantial need or inability to obtain the substantial equivalent by other means (¶¶ 21, 26); that data extracted with the tool was not testimonial, so the criminalist who ran the extraction, who followed prompts and did not analyze the data, did not have to testify (¶¶ 24, 32); and that Rule 702 and Daubert did not require exclusion or a reliability hearing where the detective and forensic examiner offered no opinions on the tool’s reliability and used different software to review the extraction (¶ 28). A petition for review was filed on August 24, 2026 (No. CR-26-0220-PR), and the Arizona Supreme Court’s docket listed it as pending as of September 2, 2026; its later status should be confirmed on the docket. The decision did not address the private-search doctrine or hash matching.
Special Issues in Legal Practice
The principles described above apply with equal force to every party in a legal proceeding. The science operates the same way regardless of who is asking the questions, and the research summarized here does not by itself favor either side. What differs between the parties is not the science but the vantage point — the particular places where these issues tend to surface, and the particular errors each party is best positioned to guard against. The two summaries below identify those recurring issues. They are offered for orientation, not as litigation strategy; in every case the controlling question is what the evidence in that case will, and will not, support, and the expert’s role is to inform the trier of fact rather than to advocate for either side.
Special Issues for Prosecutors
The research supports treating a match to a previously classified database image as strong evidence about the content of a flagged file, and independent testing found PhotoDNA highly robust to the compression and resizing images ordinarily undergo online. Hash information in a CyberTip, together with subscriber records and peer-to-peer logs, can supply substantial probable cause, and Johnsen holds that hash matches need not be confirmed by viewing the suspect’s own copies. In Arizona courts, Fristoe treats the provider as a private actor and applies the private-search doctrine under the state constitution when government review is limited to images the provider viewed. Where the law follows Wilson, Maher, and Lowers or is unsettled, prosecutors are often best positioned to avoid the issue by obtaining a warrant before anyone opens a file the report shows the provider did not view and, where reliability is contested, by presenting evidence of how the provider’s system works rather than relying on judicial notice. Good faith, inevitable discovery, attenuation, and independent source have each preserved evidence in recent cases, and Richards and Ixcolgonzalez support CSAM-limited authorizations that incorporate a detailed affidavit.
The research and cases also counsel caution. Holmes shows that the remedial exceptions can fail when an alternative-discovery theory is speculative. Affidavits that overstate what a CyberTip shows, such as describing a login address as an upload address (Armour), or that rely on generalized assertions about how people store images (Nieto), invite suppression litigation. Peer-to-peer contact-offense figures are dated, investigator-reported group estimates that do not describe the person under investigation, and without an independent production error rate for perceptual hashing, claims of near-certainty should be supported rather than asserted. Examining seized media for an offense the warrant does not name, long after the warrant offenses have been resolved, can make the search unreasonable (Gurczynski).
Special Issues for Defense
The circuit split and recent cases support careful examination of how the investigation began: whether law enforcement or NCMEC opened files no one at the provider had viewed, whether the government can prove the reliability of the hash or classifier with record evidence, and whether the affidavit accurately described what the report and subscriber records showed. Defense counsel are often best positioned to obtain the complete CyberTip, the type of hash used (cryptographic or perceptual), any threshold or classifier involved, and NCMEC’s handling of the report. The PhotoDNA testing and adversarial-hashing studies support questioning descriptions of hash matching as virtually certain, a reliability question Miller left open because the defendant had not raised it, the limits of IP attribution support scrutiny of the link between an address and a person, Holmes shows that suppression remains possible in Arizona federal cases, and Fristoe did not decide how Arizona courts treat files no one at the provider viewed.
The research and cases also counsel caution about these arguments. The robustness studies show how algorithms behave in the laboratory, not that any particular match was wrong, and a match may be confirmed by the files later found on the device. Providers are generally treated as private actors (Ringland, Bebris, Miller, Fristoe), and Fristoe applies the private-search doctrine under the Arizona Constitution; hash matches can supply probable cause without visual review (Johnsen); Mixton forecloses a constitutional challenge to IP and subscriber subpoenas in Arizona courts; and Rule 16.2 requires a specific prima facie showing before the State’s burden attaches in warrant cases. Even where a violation is found or assumed, good faith, inevitable discovery, attenuation, or independent source prevented suppression in most of the cases summarized here (Maher, Lowers, Tolbert, Armour, and the Playpen cases), and in courts-martial a guilty plea waives Fourth Amendment issues unless the accused enters a conditional plea with the military judge’s approval and the government’s consent (Mil. R. Evid. 311(e); R.C.M. 910(a)(2)).
This paper is for educational purposes only and is not legal advice. The author is a psychologist, not an attorney; summaries of law are general and must be verified against current authority or with a licensed attorney.
References
Abelson, H., Anderson, R., Bellovin, S. M., Benaloh, J., Blaze, M., Callas, J., Diffie, W., Landau, S., Neumann, P. G., Rivest, R. L., Schiller, J. I., Schneier, B., Teague, V., & Troncoso, C. (2024). Bugs in our pockets: The risks of client-side scanning. Journal of Cybersecurity, 10(1), Article tyad020. https://doi.org/10.1093/cybsec/tyad020
Bissias, G., Levine, B., Liberatore, M., Lynn, B., Moore, J., Wallach, H., & Wolak, J. (2016). Characterization of contact offenders and child exploitation material trafficking on five peer-to-peer networks. Child Abuse & Neglect, 52, 185–199. https://doi.org/10.1016/j.chiabu.2015.10.022
Farid, H. (2021). An overview of perceptual hashing. Journal of Online Trust and Safety, 1(1). https://doi.org/10.54501/jots.v1i1.24
National Center for Missing & Exploited Children. (2025, May 8). NCMEC releases new data: 2024 in numbers. https://www.missingkids.org/blog/2025/ncmec-releases-new-data-2024-in-numbers
National Center for Missing & Exploited Children. (2026a). CyberTipline data 2025. https://www.missingkids.org/cybertiplinedata
National Center for Missing & Exploited Children. (2026b, March 31). The work never stops: A first look at NCMEC’s 2025 data. https://www.missingkids.org/blog/2026/the-work-never-stops-first-look-at-ncmecs-2025-data
Seto, M. C. (2025). Online sexual offending: Theory, practice, and policy (2nd ed.). American Psychological Association. https://doi.org/10.1037/0000428-000
Steinebach, M. (2023). An analysis of PhotoDNA. In Proceedings of the 18th International Conference on Availability, Reliability and Security (pp. 1–8). Association for Computing Machinery. https://doi.org/10.1145/3600160.3605048
Steinebach, M. (2024). Robustness and collision-resistance of PhotoDNA. Journal of Cyber Security and Mobility, 13(3), 541–564. https://doi.org/10.13052/jcsm2245-1439.1339
Struppek, L., Hintersdorf, D., Neider, D., & Kersting, K. (2022). Learning to break deep perceptual hashing: The use case NeuralHash. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (pp. 58–69). Association for Computing Machinery. https://doi.org/10.1145/3531146.3533073
Wolak, J., Liberatore, M., & Levine, B. N. (2014). Measuring a year of child pornography trafficking by U.S. computers on a peer-to-peer network. Child Abuse & Neglect, 38(2), 347–356. https://doi.org/10.1016/j.chiabu.2013.10.018
Legal Authorities
Carpenter v. United States, 585 U.S. 296 (2018).
Chatrie v. United States, 609 U.S. 605, No. 25-112 (2026) (decided June 29, 2026).
State v. Eisfeldt, 185 P.3d 580 (Wash. 2008).
State v. Fristoe, 251 Ariz. 255, 489 P.3d 1200 (App. 2021) (No. 2 CA-CR 2019-0064, May 20, 2021), review denied (Ariz. Mar. 1, 2022).
State v. Ortega, No. 2 CA-CR 2024-0044 (Ariz. App. June 24, 2026) (mem. decision), petition for review filed, No. CR-26-0220-PR (Ariz. Aug. 24, 2026) (pending per the court’s docket as of Sept. 2, 2026) (cited for persuasive value only under Ariz. R. Sup. Ct. 111(c)).
Gasper v. Wisconsin, No. 25-1191 (U.S. Oct. 5, 2026) (cert. denied).
State v. Gasper, 2026 WI 3, No. 2023AP2319-CR (Wis. Jan. 14, 2026).
State v. Mixton, 250 Ariz. 282, 478 P.3d 1227 (2021).
United States v. Ackerman, 831 F.3d 1292 (10th Cir. 2016).
United States v. Ackerman, No. 17-3238 (10th Cir. Feb. 26, 2020) (unpublished).
United States v. Anzalone, 208 F. Supp. 3d 358 (D. Mass. 2016), aff’d, 923 F.3d 1 (1st Cir. 2019).
United States v. Armour, 86 M.J. 515 (A.F. Ct. Crim. App. 2025), aff’d and remanded, ___ M.J. ___, No. 26-0116/AF (C.A.A.F. Aug. 24, 2026).
United States v. Bebris, 4 F.4th 551 (7th Cir. 2021).
United States v. Blackburn, 80 M.J. 205 (C.A.A.F. 2020).
United States v. Gurczynski, 76 M.J. 381 (C.A.A.F. 2017).
United States v. Henderson, 906 F.3d 1109 (9th Cir. 2018).
United States v. Holmes, 121 F.4th 727 (9th Cir. 2024).
United States v. Horton, 863 F.3d 1041 (8th Cir. 2017).
United States v. Ixcolgonzalez, ___ M.J. ___, No. 25-0243/MC (C.A.A.F. Aug. 24, 2026).
United States v. Jacobsen, 466 U.S. 109 (1984).
United States v. Johnsen, 176 F.4th 1208 (9th Cir. 2026).
United States v. Levin, 874 F.3d 316 (1st Cir. 2017).
United States v. Lowers, 170 F.4th 134 (4th Cir. 2026).
United States v. Maher, 120 F.4th 297 (2d Cir. 2024).
United States v. McLamb, 880 F.3d 685 (4th Cir. 2018).
United States v. Miller, 982 F.3d 412 (6th Cir. 2020).
United States v. Nieto, 76 M.J. 101 (C.A.A.F. 2017).
United States v. Perkins, 78 M.J. 381 (C.A.A.F. 2019).
United States v. Reddick, 900 F.3d 636 (5th Cir. 2018).
United States v. Richards, 76 M.J. 365 (C.A.A.F. 2017).
United States v. Ringland, 966 F.3d 731 (8th Cir. 2020).
United States v. Tolbert, 92 F.4th 1265 (10th Cir. 2024).
United States v. Werdene, 883 F.3d 204 (3d Cir. 2018).
United States v. Wilson, 13 F.4th 961 (9th Cir. 2021).
United States v. Workman, 863 F.3d 1313 (10th Cir. 2017).
U.S. Const. amend. IV.
Ariz. Const. art. 2, § 8.
Revising Existing Procedures On Reporting via Technology Act (REPORT Act), Pub. L. No. 118-59, 138 Stat. 1014 (2024).
18 U.S.C. § 2258A.
Ariz. Rev. Stat. Ann. § 13-3018.
Ariz. Rev. Stat. Ann. § 13-3925.
Fed. R. Crim. P. 41(b)(6).
Ariz. R. Crim. P. 16.2.
Manual for Courts-Martial, United States, Mil. R. Evid. 311, 315 (2024 ed.).
Manual for Courts-Martial, United States, R.C.M. 703A, 910(a)(2) (2024 ed.).
Exec. Order No. 14130, 2024 Amendments to the Manual for Courts-Martial, United States, 89 Fed. Reg. 105343 (Dec. 27, 2024).
Related topic papers: Digital Evidence and What It Can Show About the User · Child Sexual Abuse Material: Federal, Arizona, and Military Law · Introduction to Online Sexual Offending · Estimating Age from Images · Online and Contact Offenders Compared
Related topics from Arizona Forensic Services: Courts-Martial