How Amazon Could use Privacy Tech to Regain Trust and Lower Prices for Customers

On the last day of August, the US Federal Trade Commission (FTC) and the attorneys general of 22 states filed a lawsuit against Amazon for engaging in "deceptive and unfair practices that secretly inflated prices in its online search advertising auctions."¹ The suit alleges that Amazon was secretly participating in the advertising auctions on its own platform in order to boost the prices of ads, making $20 billion or more in the process and having those costs passed on to smaller businesses and consumers as a result.

How Amazon's Advertising Auction System Works

When users perform a search on Amazon's website or within its mobile applications, several of the initial results are advertisements related to the search term entered by the user. Advertisers and businesses that sell products on Amazon's platform are able to place bids to compete for these ad placements via a competitive auction. Many, if not most, advertising platforms across the Internet work in a similar manner². In the case of Amazon searches and many other ad platforms, the winning bidder ends up paying whenever a user clicks on that advertisement.

In order to select the winning bid on their platform, Amazon uses a "Generalized Second Price" (GSP) system. This process primarily consists of two steps:

1. Each participant submits a bid with the maximum value they're willing to pay to show an advertisement
2. The winner is the highest bidder, but they only pay what the second-highest bidder submitted, plus $0.01.

Let's use a hypothetical auction with two participants, Bookseller A and Bookseller B, as an example, who are competing to show adds to users who search for "self-help bestseller." If Bookseller A bids for an advertisement at $2, and Bookseller B bids at $3, then Bookseller B would win but only pay $2.01, since that is one cent more than Bookseller A's second place bid of $2.

What Amazon Allegedly Did Wrong

According to documents uncovered by the FTC and state AGs and cited in their lawsuit, in multiple product categories searched for by customers, Amazon added its own auction submission to this process that would inflate the second price. They allege that Amazon did not disclose this to the other bidding advertisers³. Obviously Amazon data scientists and advertising managers would have access to more information to make a competitive offer, including large amounts of data from previous auctions that could reveal patterns in bidding behavior to derive an optimal bid and obtain the highest overall price per ad-click.

In this scenario the bidding process described above would then have an additional step:

1. Each participant submits a bid with the maximum value they're willing to pay to show an advertisement
2. Amazon uses its priviledged position to see the submitted bids and offer a new submission between the first and second bids.
3. The winner is still the highest bidder, but now they only pay what Amazon itself submitted, plus $0.01.

If the allegation turns out to be true, it's hard not to see how these costs get passed on to businesses selling on Amazon and customers buying on the platform.

Where Privacy Technologies Come In

Regardless of which way the lawsuit goes, this will be an expensive legal battle for Amazon. There are technologies and tools available that could have prevented this scenario, some of which the company already has available. These technologies involve processes that can calculate the highest value among competing bids without the organizer of the auction ever knowing who submitted them until the winner is announced. These tools are respectively known as multi-party computation, or MPC (sometimes also called "secure multi-party computation," or SMPC) and trusted-execution environments, or TEEs.

What is Multi-Party Computation (MPC)?

Multi-party computation is a way to split a calculation among multiple parties in a way that obtains the final, correct value and also hides the input values and calculation steps from any participants. In other words: multiple participants can compute a value (such as determining the highest of multiple submitted values) without needing to see each other's individual inputs.

For an online ad auction the MPC process might work like this:

1. Bookseller A submits a bid of $2, Bookseller B a bid of $6, and Bookseller C a bid of $4
2. The MPC system anonymously calculates the 2nd-highest bid
3. Bookseller B wins, paying $4.01

Crucially, in this scenario Bookseller B does not know which participant submitted a $4 bid. In fact no bidder knows the bids placed by any other bidder⁴.

Danish beet auction example

One of the first large-scale implementations of MPC came in 2008 with an auction to determine the price of sugar beets in Denmark. The association representing Danish sugar beet growers was concerned that reduced EU sugar subsidies and a reduction of Danisco's (Denmark's only sugar beets processor) production capacity would combine to give Danisco advantages in contract negotiations.

By using MPC, the farmers were able to keep their contract bids confidential from Danisco, preventing the company from potentially learning information about each farmer (such as their crop yield that year) that could reduce the size of contracts offered to individual farmers. Besides meeting the farmers' overwhelming desire for bid confidentiality, the MPC process also had the added side-effects of running more quickly and cheaply than the traditional auction process.

What is a Trusted Execution Environment (TEE)?

A modern computer processor serves numerous functions and is running many different software processes at any one time. It's often possible for a particular process (say, a software application like Word or Excel) to see the data being used by other processes. Usually this is done by design, but sometimes it can be done because of computer bugs and flaws that can be exploited by bad actors and malicious software. A Trusted Execution Environment, or TEE, addresses this risk by providing a "secure enclave" (another term sometimes used for a TEE) that isolates code and data to prevent other processes from seeing this computation taking place. Essentially, a TEE can act as a black box where the inputs and outputs are known, but what goes on inside the TEE itself is protected.

What Amazon Could do Differently

The Amazon ads team could take this opportunity to both reestablish trust and set the standard for next-generation online advertising auctions. By setting up an auctioning system that incorporates MPC and TEEs, advertising bidders would be able to anonymously submit a bid that will be evaluated fairly and without being viewed by other auction participants.

Amazon already has a strong TEE product in AWS's Nitro Enclaves that could used as part of this process. Because of the privacy and security guarantees that TEEs enable, even Amazon's own teams would not have access to MPC calculations done on Nitro Enclaves. And because an MPC process runs across multiple devices, no single Nitro Enclave would have access to the full set of bids.

By shifting to advertising auctions that utilize technologies like MPC and TEEs, Amazon could rebuild its trust with advertisers and demonstrate a powerful integration of privacy and security technologies. If AWS is soliciting names for a product that combines TEEs and MPC, "Nitro Network" does have a nice ring to it.