Skip to content

Latest commit

 

History

History
363 lines (290 loc) · 19 KB

FLEDGE_extended_PA_reporting.md

File metadata and controls

363 lines (290 loc) · 19 KB
Raw

FLEDGE has been renamed to Protected Audience API. To learn more about the name change, see the blog post

Extended Private Aggregation Reporting in FLEDGE

Introduction

The Private Aggregation API proposes a mechanism for participants of FLEDGE auctions to create summary reports which can measure generic cross-site data.

This proposal seeks to extend this API to support further measurement use-cases for FLEDGE auctions using the same mechanism to aggregate reports from the browser. The new APIs added allow for reporting bidding information and new signals (post-auction signals, latency) at different points in auction (e.g. win, loss, bid, click), while retaining the same privacy and aggregation aspects of the Private Aggregation API proposal.

Goals

  • Allow all auction participants (both winner and losers) to measure aggregate data about their bids, competing bids, and other auction signals
    • “How far off the auction clearing price was my bid?”
    • "Why was my bid rejected?"
    • "How far off the second highest bid value was my bid?"
  • Allow auction participants to measure aggregate data using signals only available during bid generation
    • "How many times did my campaign bid?"
    • "What is my campaign's win rate?"
  • Allow auction winners to measure aggregate data about the auction relative to events which happen with the rendered ad (such as a click on the ad)
    • "What is the bid cost per click?"
    • "What is the bid cost per render?"
  • Allow auction sellers to measure aggregate data on auction latency and other statistics
    • "Which participants are contributing the most to auction latency?"
  • Allow auction buyers to measure aggregate data on their bidding latency
    • “How much latency are my bidding signals fetches and bid generation code contributing?

Background

The Private Aggregation API currently allows auction winners to generate histogram contributions within their reportResult()/reportWin() methods using a new API available in the worklet:

function reportResult(auctionConfig, browserSignals) {
  …
  privateAggregation.sendHistogramReport({
      bucket: convertBuyerToBucketId(browserSignals.interestGroupOwner),
      value: convertBidToReportingValue(browserSignals.bid)
    });
}

While this provides a way to gather important information about a winning creative, for instance, the average bid submitted for a creative, it still lacks some information that can be useful for auction participants as well as sellers. Information which is also not available through event level reporting. Just to mention a few:

  • The ability for a seller to track the average number of auction participants
  • The ability for a buyer to know the win rate of a campaign
  • The ability for a buyer/seller to know the highest other bid (HOB) for training bid shading models
  • The ability for a buyer/seller to measure actions contributing to auction latency
  • The ability for a buyer to link useful bidding signals (for instance time a user has spent in an interest group) with a click event — this information is essential for model training.

Below, we propose an extension of the Private Aggregation API that tries to address these use cases.

Declaring Contributions

A natural way to address the above use cases is for auction participants and sellers to be able to trigger arbitrary reports at different events. For instance, a buyer may want to count the number of times users that were in an interest group for less than an hour clicked on the winning ad. Or they may want to track auction signals associated with a non-winning bid, etc.

Our proposed API extension allows a bidder to register a set of histogram contributions associated with an arbitrary event_key within generateBid, scoreAd, reportWin, and reportResult.

Example 1: Correlating bidding signals with click information.

We consider the scenario where a buyer wants to learn the click through rate of ads when a user has been in an interest group for a given time.

The buyer may implement getImpressionReportBucket() and getClickReportBucket() which map an interest group and the time the user has spent in that interest group to a 128-bit integer.

The buyer can then do the following during generateBid (when the above information is available)

function generateBid(interestGroup, auctionSignals, perBuyerSignals, trustedBiddingSignals, browserSignals) {
 …
  privateAggregation.reportContributionForEvent(“reserved.win”, {
      bucket: getImpressionReportBucket(),
      value: 1
  });
  privateAggregation.reportContributionForEvent("click", {
      bucket: getClickReportBuckets(), // 128-bit integer as BigInt
      value: 1
  });

The above logic will trigger a report if the generated bid wins (see reserved.win). And another one, if the user later clicks on the winning ad (this needs to be triggered by the fenced frame itself, see reportPrivateAggregationEvent. When the buyer receives an aggregated report they can infer what the click-through-rate (CTR) was for users on different “interest group age” buckets.

Example 2: Getting the average bid gap for an ad.

A buyer may be interested in understanding how much higher they should have bid in order to win an auction. To do this, not only do we need to trigger a report at loss time (see Reporting for bids which do not win), but also have the value of the report depend on the outcome of the auction. To do this, we introduce a field to the “contributions” object called signalValue. This field allows the report to depend on post auction information. A signalValue object is composed of the following values:

  • baseValue: The name of the auction result value we want to report. For instance, winning-bid.
  • scale: Optional scale factor by which we want to multiply the auction result value. This is useful for controlling the amount of noise added by the aggregation service. Scale is applied before offset is added.
  • offset: Optional offset to add to the auction result value.

After the auction happens, the final value of the generated report is (baseValue * scale) + offset. The following example shows how to return the gap between an ad bid and the winning bid:

function generateBid(...) {
  bid = 100;
  privateAggregation.reportContributionForEvent(
    "reserved.loss",
    {
      bucket: 1596n, // represents a bucket for interest group x winning bid price
      value: {
        baseValue: "winning-bid",
        scale: 2, // Number which will be multiplied by browser value
        offset: -bid * 2 // Numbers which will be added to browser value, after scaling
       }
    });

  return bid;
}

In the event this bid does not win the auction, and the winning bid was 200, the histogram contribution would be generated as:

bucket: 1596n
value: 200 // = (winning-bid * scale) + offset = (200 * 2) + (-100 * 2)

This would correspond to the gap by which the advertiser lost the auction, scaled by 2. Scaling is important as it allows bidders to better control the amount of noise which will be added to various aggregation keys.

Example 3: Understand the reason an ad did not win

Similar to the above example, sometimes, the key that we want to aggregate over may depend on the outcome of an auction. To solve this use case we provide an object called signalBucket. The final bucket id of the bucket will depend on the outcome of the auction. The following example allows the buyer to keep track of how many times their bid was rejected for particular reasons.

function generateBid(...) {
  privateAggregation.reportContributionForEvent(
    "reserved.loss",
    {
      bucket: {
        baseValue: "bid-reject-reason",
        offset: 500n // Offset buckets
       },
      value: 1
    });

  return bid;
}

If the bid is rejected for being below auction floor, this would result in a contribution being generated:

bucket: 502n // 500n + 2n (2n is the value associated with bids below auction floor, see bid-reject-reason below)
value: 1

Reporting API informal specification

privateAggregation.reportContributionForEvent(eventType, contribution)

The parameters consist of:

  • an eventType which is a string identifying the event type that triggers this report to be sent (see Triggering reports below), and
  • a contribution object which contains:
    • a bucket which is a 128bit ID or a signalBucket which tells the browser how to calculate the bucket (represented as BigInt) and
    • a value which is a non-negative integer or a signalValue which tells the browser how to calculate the value.

Where signalBucket and signalValue is a dictionary which consists of:

  • a baseValue field indicating which value the browser should use to calculate the resulting bucket or value. A signalValue.baseValue or signalBucket.baseValue may be any of the following:
    • winning-bid: the value used is the winning bid value.
    • highest-scoring-other-bid: the value used is the bid value that was scored as second highest.
    • script-run-time: milliseconds of CPU time that the calling function required, when called.
    • signals-fetch-time: milliseconds required to fetch the trusted bidding or scoring signals, when called from generateBid() or scoreAd() respectively.
    • bid-reject-reason: one of the following values:
      • 0: indicates seller rejected bid without providing a reason, i.e., bid reject reason not available
      • 1: indicates seller rejected bid because “Invalid Bid”
      • 2: indicates seller rejected bid because “Bid was Below Auction Floor”
      • 3: indicates seller rejected bid because “Creative Filtered - Pending Approval by Exchange”
      • 4: indicates seller rejected bid because “Creative Filtered - Disapproved by Exchange”
      • 5: indicates seller rejected bid because “Creative Filtered - Blocked by Publisher”
      • 6: indicates seller rejected bid because “Creative Filtered - Language Exclusions”
      • 7: indicates seller rejected bid because “Creative Filtered - Category Exclusions”
      • 8: indicates seller rejected bid because "Creative Filtered - Did Not Meet The K-anonymity Threshold"
      • Perhaps other values indicating:
        • generateBid() hitting timeout
        • The auction was aborted (i.e. calling endAdAuction())
        • an auction that never rendered the ad
  • optional offset and scale that allow the reporter to manipulate the browser signal to customize the buckets and values they will receive in reports:
    • scale will be multiplied by the browser provided value. Scale is applied before offset is added. Default value is 1.0.
    • offset will be added to the browser provided value (allows you to shift buckets, etc). Default value is 0.

Triggering reports

Reporting bidding data associated with an event in a frame

A fenced frame can trigger the sending of contributions associated with an arbitrary event by calling into a new API:

window.fence.reportEvent("click");

This will cause any contributions associated with a call to reportContributionForEvent() with an event-type of click to be reported/sent.

In this example, "click" is an event-name chosen by the auction bidder. There are a number of event names that are reserved and invoked directly by the browser. All reserved values will have the "reserved." prefix, and all non-reserved values cannot have the "reserved." prefix.

Reporting bidding data for wins

The reserved.win event-type is a special value that indicates a report should be sent when a bid wins the auction.

The browser will trigger sending if contributions were registered for a winning bid.

Reporting for bids which do not win

The reserved.loss event-type is a special value that indicates a report should be sent when a bid does not win the auction. If the bid does not win, the browser will trigger reporting for the loss contributions.

Reporting bidding data no matter whether wins or not

The reserved.always event-type is a special value that indicates a report should be sent no matter whether a bid wins the auction or not. The browser will always trigger reporting for the always contributions.

Reporting Per-Buyer Latency and Statistics to the Seller

The seller may want to collect aggregate statistics on latency and bids for their auctions. Measuring this per buyer will allow the seller to evaluate performance information at a more granular level and give helpful guidance and feedback to their bidders.

In order for the seller to collect this additional information from buyers, we need to give buyers a way to express which sellers they’re comfortable sharing this information with, so we add a new mechanism which allows each buyer to declare a set of approved sellers: The interestGroup provided to navigator.joinAdInterestGroup() will contain a new field named sellerCapabilities, a dict keyed by seller origin (or "*", to set defaults for non-specified seller origins) with lists of permission strings as values (as described below). Sellers might want to only score bids from interest groups that will share aggregate statistics, so a field, requiredSellerCapabilities will also be added to the auction config. Any interest group that doesn't permit (for the auction's seller) all the sellerCapabilities listed in requiredSellerCapabilities will not participate in the auction.

For the seller to declare reporting, the auctionConfig passed to runAdAuction is amended to contain a configuration for the seller latency report.

const auctionConfig = {
  'seller': 'https://www.example-ssp.com',
  …
  ‘interestGroupBuyers’: ['https://buyer1.com', 'https://buyer2.com', …],
  // An aggregation key for each buyer. This represents the starting contribution bucket
  // associated with the corresponding buyer.
  ‘auctionReportBuyerKeys’: [100n, // key for buyer1.com
                             105n, // key for buyer2.com
                             …],
  // Configures what values will be reported for each buyer. The key declares what signal will be
  // measured, and the bucket declares the bucket this will be placed relative to the
  // buyer key and the scale determines how the value is scaled.
  `auctionReportBuyers`: {
    `interestGroupCount`:       { `bucket`: 0n, `scale`: 1,  },
    ‘bidCount’:                 { `bucket`: 1n, `scale`: 1,  },
    ‘totalGenerateBidLatency’:  { `bucket`: 2n, `scale`: 1,  },
    ‘totalSignalsFetchLatency’: { `bucket`: 3n, `scale`: .1, },
  }
}

The seller is able to measure the following for each buyer, assuming permission is granted via the indicated sellerCapabilities for that seller:

  • interestGroupCount: The number of the interest groups which could participate in the auction (i.e. the number of intererest groups on the machine for this buyer -- note the count isn't limited by the auction config's perBuyerGroupLimits). This requires the interest-group-counts sellerCapabilities permission.
  • bidCount: The number of valid bids generated by this buyer. This requires the interest-group-counts sellerCapabilities permission.
  • totalGenerateBidLatency: The sum of execution time for all generateBids() in milliseconds. This requires the latency-stats sellerCapabilities permission.
  • totalSignalsFetchLatency: The total time spent fetching trusted buyer signals in milliseconds. If the interest group didn't fetch any trusted signals, then 0 milliseconds is reported. This requires the latency-stats sellerCapabilities permission.

Given the auctionConfig above, if buyer1.com had two interest groups participate in the auction, their trusted buyer signals fetch taking 10ms, their generateBid() scripts running for 2ms and 3ms, and one of the interest groups returning a bid, the following histogram reports would be generated by the browser:

{ bucket: 100n,// = 100n + 0n = buyer1.com’s key + interestGroupCount’s bucket
  value: 2 }, // = 2 interest groups were given the chance to bid.
{ bucket: 101n,// = 100n + 1n = buyer1.com’s key + bidCount’s bucket
  value: 1 }, // = 1 interest group returning a bid.
{ bucket: 102n,// = 100n + 2n = buyer1.com’s key + totalGenerateBidLatency’s bucket
  value: 5 }, // = 3ms + 2ms generateBid() runtimes.
{ bucket: 103n,// = 100n + 3n = buyer1.com’s key + totalSignalsFetchLatency’s bucket
  value: 1 }, // = 10ms total trusted bidding signals fetch time multiplied by .1 scale.

Here's an example of what the sellerCapabilities interest group field could look like:

'sellerCapabilities': {
  'https://seller.com': [ 'interest-group-counts', 'latency-stats' ],
  'https://seller2.com': [ 'latency-stats' ],
  '*': [ 'interest-group-counts' ]
}

This would grant both interest-group-counts and latency-stats permission to https://seller.com, latency-stats to https://seller2.com, and interest-group-counts to all other sellers.

NOTE: the permission names interestGroupCounts and latencyStats are deprecated and will be removed in a future Chrome release, as they do not follow the documented WebIDL naming conventions.

Data Volume

To control the data volume of aggregatable reports, auction participants may want to subsample reports on the client to avoid high costs or latency associated with aggregation. Auction participants can do this by only utilizing the Private Aggregation API some fraction of the time.

Subsampling on the client side as opposed to the server is ideal to avoid oversampling infrequent events / campaigns.

Privacy Considerations

See Private Aggregation Privacy Consideration

Reports generated with the mechanisms in this proposal will be subject to the same contribution bounding and bucketing for Private Aggregation Reports.

While new data is being used to generate histogram contributions (bidding signals, information from within the fenced frame), the general privacy properties offered by the Aggregation Service still hold.

Ideas For Future Iteration

Including remaining user contribution budget as an input to reports

In the current proposal, the browser limits the overall sensitivity of the contributions made by a single user per reporter per day. It may be desirable to allow reporters/interest group owners to select their histogram contributions knowing the amount of budget which remains. Currently, reports created after the budget has been exhausted are ignored.

Reporting bidding data for renders

We may want to add a reserved.render event-type or similar which allows a bidder to measure contributions when the browser shows the ad, or potentially when the browser finishes fetching the document.