Skip to content

Latest commit

 

History

History
457 lines (360 loc) · 22 KB

report_verification.md

File metadata and controls

457 lines (360 loc) · 22 KB

Preventing invalid aggregatable reports via report verification

Table of Contents

This document describes a basic integration between the Attribution Reporting API and Private State Tokens API, focused on aggregatable reports. The purpose of this integration is to enhance the security of the aggregatable reports by making it more difficult for bad actors to interfere with the accuracy of attribution measurement. This integration is possible via a new flow where reporting origins can submit private tokens attesting to the veracity of a trigger registration.

Note: many of the ideas in this document are also compatible with the Private Aggregation API.

Security goals

The primary security goal of this proposal is to prevent a class of attacks on forging fake aggregatable reports, for reporting origins that opt into the feature. The attacker we are concerned about is a party attempting to falsify reports, to corrupt aggregate measurement.

Our security goals for this proposal are as follows:

  • No reports out of thin air: It should not be possible for an attacker to generate fake reports out of thin air.

  • No replaying reports: It should not be possible for an attacker to replay one honest report multiple times, if they can exfiltrate an honest report.

We also have a few secondary goals we’d hope the solution achieves:

  • Privacy of the invalid traffic (IVT) detector: Ideally, a client should not be able to use the output of this system to learn anything about the reporting origin’s invalid traffic detector (and therefore, learn how to bypass it more easily). In other words, the output of the IVT detector should be private to the client.

  • Limit the attack scope for bad actors that can bypass IVT detectors: The proposal leverages existing IVT detectors operated by the reporting endpoint. Any proposal should attempt to limit the attack scope even for adversaries that can craft requests to bypass detection.

  • No report mutation: It should not be possible for an attacker to mutate reports arbitrarily after they are generated by an honest client. Note this is only P2 because the API only works over HTTPS, so mutating reports from an honest client requires a very specific attacker with access to infrastructure owned by the reporting origin (or something like "read-only" malicious client malware).

Existing mitigations and practical threats

"No reports out of thin air" is not currently prevented in the API, without the proposal in this document. However, practical attacks on the system are challenging, as they require an adversary to have a deep understanding of the reporting origin’s key space in order to alter aggregate results. A much more practical attack is one closer to replaying honest reports, since it is fairly easy for an adversary to learn a set of honest reports via instrumented web browsing, and to replicate them.

Blind signatures overview

At a high level, Private State Tokens implements a blind signature-style operation:

BlindMessage := Blind(message)

BlindSignature := Sign(blindedMessage)

ClearSignature := Unblind(blindSignature)

Verify(message, ClearSignature) => boolean

In this document, we use the following terms:

  • Blinded token: Equivalent to the BlindMessage
  • Token: Equivalent to the (message, ClearSignature) pair

For simplicity, this omits the public/private metadata aspects of Private State Tokens which extend this basic primitive.

Reporting origin opt-in

In order for a reporting origin to use this feature, they must configure themselves as a Private State Token issuer, by publishing key commitments publicly (see issue 117 on the Private State Token repo, as this process is still being finalized).

API changes

sequenceDiagram
    participant B as Browser
    participant A as Adtech
    B->>A: Source ping
    A->>B: Source registration
    Note over A,B: Time passes
    B->>A: Trigger ping, bearing blinded id/dest
    A->>A: Check IVT filter
    A->>B: Trigger registration, with blind-signed id/dest
    B->>B: Verify signature and unblind id/dest + signature
    Note over A,B: Time passes
    B->>A: Send report alongside signature
Loading

Issuance

The token issuance protocol can only happen on trigger registration. For reporting origins that are configured to do Private State Token issuance, all new requests that have the Attribution-Reporting-Eligible request header with the trigger value will automatically attach a Sec-Attribution-Reporting-Private-State-Token header with a comma-separated list of blinded messages (which embeds the concatenation of the report_id and attribution_destination).

The reporting origin will inspect the request and decide whether it is valid, i.e. whether the origin suspects it is coming from a real, honest client and should therefore be allowed to generate attribution reports.

  • If the request is considered invalid and hence shouldn’t be taken into account to calculate aggregate measurement results, the reporting origin should respond without adding a Sec-Attribution-Reporting-Private-State-Token response header. If this header is omitted or is not valid, the browser will proceed with trigger registration normally, but any report generated will not contain the verification header. Note: more advanced cases deployments can consider issuing an "invalid" token using private metadata to avoid the client learning the detection result. See privacy of the IVT detector for more details.
  • If the request is considered valid, the reporting origin should add a Sec-Attribution-Reporting-Private-State-Token header with blind tokens (the blind signatures over the blinded messages).

Internally, the browser will randomly associate the tokens with any reports generated from this trigger request, including null reports.

Note: because we only support issuance requests on requests bearing the Attribution-Reporting-Eligible header, some trigger requests flows are not supported due to issue #347. Supporting those flows would require performing the issuance protocol on almost every resource request on the web.

Redemption

Any tokens associated with an attribution report will be sent along with the report’s request in the form of a new request header Sec-Attribution-Reporting-Private-State-Token. If this token successfully verifies, then the reporting origin has a guarantee that the report was associated with a previous trigger request that was signed.

Note: unlike the basic Private State Token API (which enables conveying tokens from one site to another), there are no redemption limits for Attribution Reporting API integration. See privacy considerations for more information.

Privacy considerations

The privacy goal of this integration is to be as privacy-neutral as possible. For this, we want to ensure that token integration doesn’t add any cross-site information leakage to the API. In particular, the worry here is that the encrypted reports can be partitioned by extra data that the tokens (or lack thereof) communicate across sites. If the count of encrypted records is sensitive, then partitioning this count per token-state (trigger-side data) could amplify a counting attack.

Even without report verification, reports can already be partitioned by the following trigger-side fields:

  • scheduled_report_time
  • attribution_destination

Additionally the designs to protect the number of encrypted reports attempt to neutralize the privacy harm that is possible by counting encrypted reports. At the most extreme end we can make the count of encrypted records totally public knowledge, which makes even arbitrary partitioning privacy neutral. However, with less extreme techniques there are privacy benefits to reducing the partitioning available.

For this reason, partitioning mitigations will not be implemented and the basic Private State Token protocol will be allowed with no extra restrictions. This allows the reporting origin to partition encrypted reports based on the following information:

  • Presence or absence of a token on a report leaking information about a conversion.

  • "Public/private metadata" (in the form of which public key is used to sign the token) leaking non-noisy information. Currently, the Private State Tokens API allows six buckets of metadata.

While we don’t currently plan to mitigate against this partitioning, it could in theory be mitigated by the following techniques:

  • The browser could refuse to register conversions for triggers which did not come with an issued token.

  • The browser could accept token issuance from only a single key at a time (and forbid private metadata). Note that this conflicts with the security goal of IVT detector privacy.

Security considerations

Evaluating the top level security goals:

No reports out of thin air

This is addressed because any report created out of thin air and sent to the ad-tech will need to be associated with a token. Any report not associated with a token can be immediately discarded as invalid. Tokens cannot be generated "out of thin air", because receiving one requires an interaction with the reporting origin.

No replaying reports

This is addressed because every report will come with a token bound to its report_id. Reporting origins can easily protect themselves from replays by keeping track of which IDs they have already seen (something we already recommend due to browser retries).

Privacy of the IVT detector

This is addressed via supporting the existing Private State Token's notion of Private metadata. Reporting origins may not wish to reveal to clients (who could be potential attackers) the signal of whether the report was deemed invalid or not, since this exposes information about the reporting origin's potentially secret logic. Private metadata solves this problem by embedding information that is only readable to the reporting origin that the client can merely read the cardinality of (e.g. that the information is just 1 bit). A reporting origin wishing to hide information about its IVT detector can simply issue tokens unconditionally, and embed the IVT decision in the private metadata bit.

Limit the attack scope for bad actors that can bypass IVT detectors

This is not addressed in this proposal, but we are exploring future extensions to address it.

No report mutation

This is only partially addressed via binding the issued token to the report’s report_id and attribution_destination. This means that tokens are not "fungible" and could only be used with a report with those matching properties. Note that it may be possible to ensure total immutability of a report (by signing over the whole generated report), but it naturally implies that we’d need to add another round of interaction between client and reporting origin, which may not be worth the extra complexity to support this.

Note that because we expect reporting origins to already deduplicate based on report_id, this infrastructure should neatly work to handle the "double spend" problem where attackers attempt to use tokens multiple times.

Future Extension: scalable public metadata for destination verification

Currently, Private State Tokens supports a notion of "public metadata" that corresponds essentially to which signing key is used to generate a blind token. This is difficult to scale for large domains like all possible destination sites, but we can consider adopting cryptographic extensions (e.g. 1, 2) which support signing over some public domain (e.g. sites).

Essentially, we can allow the token issuance operation to receive and sign over the destination site in cleartext instead of blind signing. This does not regress privacy, but would protect against a malicious client that is forging blinded tokens that intentionally falsify the destination site.

This allows us to support the "Limit the attack scope for bad actors that can bypass IVT detectors" goal, because a bad actor that can bypass an IVT detector but does not actually trigger any third party conversions can only submit reports constrained by which destination site they actually advertised when they were issued tokens.

FAQ

Why doesn’t this support issuance at source registration time?

The proposed designs for protecting encrypted record counts prevent us from sending data from the client that reveals non-noisy cross-site data. They do this by allowing the browser to send "null" reports that don’t affect aggregate output at all. If we introduce a mechanism where a report can attest to being part of a valid source and trigger, it prevents the browser from being able to send these null reports.

This constrains us to two solutions:

  1. Attested sources can send null reports (with the associated source token) proportional to the number of sources.
  2. Attested triggers and send null reports (with the associated trigger token) proportional to the number of triggers

This design chooses (2) because:

  • We expect triggers to have more IVT signals (i.e. they can often be accompanied by verifiable user actions like purchases)
  • We expect the volume of triggers to be less than sources, which helps with system health of the API

Why doesn’t this support event-level reports?

The reason is very similar to why this proposal doesn’t support issuance at source registration time. Event-level reports are subject to browser randomization in which the browser needs to be free to create fake triggers. This means that we couldn’t possibly provide verification for triggers as it breaks the privacy mechanism. It is possible to attest sources, but because event-level reports already have a high entropy (hard to guess) ID associated with it, the added benefit of tokens is minimal.

Why deviate from the Private Click Measurement approach?

Private Click Measurement (PCM) supports a similar design, where source and trigger registration support adding an optional nonce, which is signed via an interaction with the server. For the privacy reasons described above, we can’t align with PCM on source signing.

For trigger signing, our approaches are broadly compatible, except that

  • PCM does all configuration via query parameters in a (blocked) redirect and we do everything in headers. This is necessary for us because we support additional registrations with real redirects.
  • PCM puts the resulting unlinkable token in the body of a report, while we propose to do it in a separate header. This is mostly to align with our header issuance protocol to keep things consistent.
  • PCM uses RSABSSA while this proposal uses Private State Tokens / Privacy Pass for underlying crypto operations which sits one layer higher in the stack. Notably, Privacy Pass internally uses RSABSSA. Integrating with Private State Tokens allows us to re-use its key commitment infrastructure.

Other than that the designs are very similar.

Could we just tag reports with a trigger_id instead of using anonymous tokens?

In principle, we could abandon the token-based approach and tag every report with an unencrypted high-entropy identifier that is configured at trigger time. This would uniquely tie every encrypted report with the trigger context that generated it. Because the ids are high entropy (and thus difficult to guess), they could serve the purpose of the anonymous tokens and additionally allow for offline filtering of reports (i.e. deferring the IVT decision until after trigger registration).

This will allow a reporting origin to partition encrypted reports on a per-advertiser-user basis, which means that in order to make this idea privacy neutral, we’d need to fully protect the counts of encrypted reports e.g. by unconditionally sending aggregatable reports on every trigger registration (even when there is no corresponding source), ensuring that the encrypted payloads reveal no cross-site information. See issue #439 for more details. We are not proceeding with this option currently due to performance concerns (bandwidth, compute) with sending lots of null reports.

Note that this approach would also mean we could likely remove some/all of the delay in aggregatable reports.

How can I compose this with other trust signals?

This API allows the reporting origin to make a token issuance decision based on the information embedded in the trigger registration request. To aid in this decision it’s expected that the request will come bearing some context that will help the origin make this decision, and the proposal is extensible to support adding many forms of context in the registration request. Here are two examples:

Basic Private State Tokens

When making an "IVT" decision on trigger registration requests, it may be useful to have the extra trust signals from basic Private State Tokens. In many cases it should be straightforward to ensure that trigger registration requests come bearing Private State Tokens:

window.fetch('https://adtech.example/trigger-attribution?convid=123', {
  attributionReporting: { triggerEligible: true },
  keepalive: true,
  privateToken: {
    operation: 'send-redemption-record',
    issuers: ['https://invalid-detection.example/'],
  },
});

Custom trust signal

Any other custom trust signals available can be used as an input into the token issuance protocol. The only requirement is that the signal is exposed in the network request that is used to trigger attribution. Note that if the request comes bearing a cookie, the cookie can also be used as input into the issuance protocol.

const trustSignal = generateTrustSignal();
window.fetch(`https://adtech.example/trigger-attribution?convid=123?trustSignal=${trustSignal}`, {
  attributionReporting: { triggerEligible: true },
  keepalive: true,
});

How should this work with cross app and web integration?

Our designs for cross app and web measurement allow routing attribution to the underlying platform. Therefore, to support this kind of token-based verification in that setting would require the underlying platform to support this design.