AWS Security ChangesHomeSearch

AWS wellarchitected: Rewrite ADVCOST05-BP01 for fraud detection pipeline costs

Service: wellarchitected · 2026-11-11 · Documentation medium

File: wellarchitected/latest/video-streaming-advertising-lens/advcost05-bp01.md · Type: network

Summary

Completely replaces the previous best practice (cost-efficient data types for collaborative data environments / AWS Clean Rooms, parquet, SQL) with a new best practice on designing fraud detection pipelines: progressive filtering funnel (AWS WAF Bot Control at the edge, ElastiCache verdict caching, rule-based heuristics, ML inference only for ambiguous traffic), Spot training via SageMaker Managed Spot Training, Cost Anomaly Detection alarms, and new related documents/services links including Amazon GuardDuty.

Security assessment

No CVE, vulnerability, or incident is referenced; the change is a cost-optimization lens rewrite. However, it does document security controls and practices (AWS WAF Bot Control managed rules in Count then Block mode, verdict caching of known-bad traffic, edge filtering before application processing, and a GuardDuty reference), which is security guidance framed around cost efficiency, so it is security-adjacent rather than fixing a specific weakness.

Evidence

  1. **Implement edge-level bot detection using AWS WAF Bot Control:** Deploy WAF Bot Control managed rules on bid request endpoints. Configure rate-based rules per SSP IP range. Start in Count mode to assess traffic composition, then transition to Block mode for clearly fraudulent traffic categories. This avoids compute costs for obvious threats before they reach your application.

Diff

diff --git a/wellarchitected/latest/video-streaming-advertising-lens/advcost05-bp01.md b/wellarchitected/latest/video-streaming-advertising-lens/advcost05-bp01.md
index 8523493..49b5a6f 100644
--- a//wellarchitected/latest/video-streaming-advertising-lens/advcost05-bp01.md
+++ b//wellarchitected/latest/video-streaming-advertising-lens/advcost05-bp01.md
@@ -7 +7 @@
-Implementation guidanceKey AWS servicesResources
+Implementation guidanceResources
@@ -9 +9 @@ Implementation guidanceKey AWS servicesResources
-# ADVCOST05-BP01 Use cost efficient data types and configurations for collaborative data environments
+# ADVCOST05-BP01 Design fraud detection pipelines to minimize redundant processing and optimize inference costs
@@ -11 +11,34 @@ Implementation guidanceKey AWS servicesResources
-Use efficient storage formats and streamlined query configurations to reduce unnecessary data scanning, duplication, and transfer costs in collaborative analytics environments. 
+Design fraud detection workflows that avoid repeated evaluations. Cache known outcomes, filter threats as early as possible, and run only the necessary inference steps on cost-efficient compute resources. 
+
+**Desired outcome:**
+
+  * Your fraud detection pipeline processes each unique signal only once, caches the result for repeated encounters, and applies progressively more expensive evaluation steps only to traffic that passes cheaper upstream filters. 
+
+  * Known-bad traffic is blocked at the edge, known-good traffic bypasses deep inspection, and ML inference runs only on ambiguous cases. 
+
+
+
+
+**Common anti-patterns:**
+
+  * Running full ML inference on every bid request without first checking whether the creative, IP, or device ID has already been evaluated and cached. 
+
+  * Implementing filtering logic only at the application layer rather than at the network layer, consuming compute for traffic that could be filtered out upstream. 
+
+  * Training fraud detection models on On-Demand instances when the workload is interruptible and suitable for Spot pricing. 
+
+
+
+
+**Benefits of establishing this best practice:**
+
+  * Lower inference costs from avoiding repeated evaluation of previously seen inputs. 
+
+  * Lower overall compute spend by blocking threats at the cheapest layer before they reach expensive processing stages. 
+
+  * Early detection of cost anomalies caused by fraud attacks driving up evaluation volume. 
+
+
+
+
+**Level of risk exposed if this best practice is not established:** High 
@@ -15 +48,3 @@ Use efficient storage formats and streamlined query configurations to reduce unn
-  * Use parquet or columnar formats with partitioning and compress datasets. 
+A threat actor pays almost nothing to generate fraudulent traffic, while your cost to evaluate each request scales linearly with volume. A flat architecture that gives every request the same evaluation depth becomes economically unsustainable during attack surges. Progressive filtering orders evaluation stages by increasing cost, so cheap upstream filters remove traffic before it reaches expensive downstream stages. 
+
+The progressive funnel pattern assigns each detection technique to the cheapest layer that can run it. Network-layer rules at the edge ([AWS WAF](https://docs.aws.amazon.com/waf/latest/developerguide/what-is-aws-waf.html)) cost fractions of a cent per million evaluations because they execute on infrastructure already in the request path. Cached verdict lookups ([Amazon ElastiCache](https://docs.aws.amazon.com/AmazonElastiCache/latest/red-ug/WhatIs.html)) add microseconds of latency and sub-cent costs per lookup. Rule-based heuristics require application compute but avoid ML inference overhead. ML inference is the most expensive stage, costing orders of magnitude more per evaluation than any upstream layer. Run inference only on cases the cheaper stages cannot resolve. 
@@ -17 +52 @@ Use efficient storage formats and streamlined query configurations to reduce unn
-  * Use standard SQL for lightweight or well-partitioned datasets. 
+Cache effectiveness depends on how repetitive your traffic is. Advertising fraud often reuses the same creative IDs, device fingerprints, or IP ranges across millions of requests. A verdict cache converts the first evaluation of each unique input into an amortized cost across all subsequent encounters. Cache hit rates above 80% are typical for stable creative and device populations. Caches also go stale. If a previously legitimate creative becomes compromised after you cache it, the clean verdict persists until the TTL expires. Choose the TTL to balance cost savings (longer TTL, higher hit rate) against detection freshness (shorter TTL, faster response to newly compromised inputs). 
@@ -19 +54 @@ Use efficient storage formats and streamlined query configurations to reduce unn
-  * Avoid unnecessary cross-joins or full table scans. 
+Fraud attacks cause cost spikes because they inflate evaluation volume unpredictably. Without anomaly detection at the pipeline level, a sudden traffic surge can generate thousands of dollars in inference costs before you detect it. [AWS Cost Anomaly Detection](https://docs.aws.amazon.com/cost-management/latest/userguide/manage-ad.html) operates on historical spending patterns and reports deviations, but it has detection lag. Teams that rely only on billing alerts might discover cost spikes hours after they begin. 
@@ -21 +56 @@ Use efficient storage formats and streamlined query configurations to reduce unn
-  * Use same-Region AWS Clean Rooms collaborations to minimize inter-Region transfer costs. 
+Model training workloads tolerate interruption. Inference does not. A training job that loses its instance mid-epoch can resume from a checkpoint without losing prior computation. This makes training eligible for [Spot pricing](https://docs.aws.amazon.com/sagemaker/latest/dg/model-managed-spot-training.html) at 60 to 90 percent lower cost than On-Demand. Spot Instances can be interrupted and delayed, so training schedules must tolerate non-deterministic duration. For fraud model retraining cadences of weekly or longer, that variability rarely matters. 
@@ -22,0 +58 @@ Use efficient storage formats and streamlined query configurations to reduce unn
+### Implementation steps
@@ -23,0 +60 @@ Use efficient storage formats and streamlined query configurations to reduce unn
+  1. **Implement edge-level bot detection using AWS WAF Bot Control:** Deploy WAF Bot Control managed rules on bid request endpoints. Configure rate-based rules per SSP IP range. Start in Count mode to assess traffic composition, then transition to Block mode for clearly fraudulent traffic categories. This avoids compute costs for obvious threats before they reach your application. 
@@ -24,0 +62 @@ Use efficient storage formats and streamlined query configurations to reduce unn
+  2. **Deploy caching for fraud evaluation results:** Implement a cache-aside pattern using ElastiCache Redis keyed on creative ID, source IP, and device ID fingerprint. Before invoking ML inference, check the cache: if the input was evaluated within the TTL window, return the cached verdict immediately. Target 80%+ cache hit rate for stable creative and device populations. 
@@ -26 +64 @@ Use efficient storage formats and streamlined query configurations to reduce unn
-## Key AWS services
+  3. **Structure the fraud detection pipeline as a progressive funnel:** Configure processing stages in order: WAF rules (block known-bad at edge) → cached lookups (return known verdicts) → lightweight rule-based checks (known patterns) → ML inference (for genuinely ambiguous cases only). Measure what percentage of traffic reaches each stage. High volumes reaching ML inference may indicate overly permissive upstream filters. 
@@ -28 +66 @@ Use efficient storage formats and streamlined query configurations to reduce unn
-  * AWS Clean Rooms 
+  4. **Use Spot Instances for fraud model training:** Configure SageMaker AI training jobs with Managed Spot Training enabled. Implement checkpointing so interrupted training resumes from the last checkpoint rather than restarting. Configure alarms on Cost Anomaly Detection for the SageMaker AI training service to detect unexpected cost spikes. 
@@ -35 +73,16 @@ Use efficient storage formats and streamlined query configurations to reduce unn
-  * [Data formats for AWS Clean Rooms](https://docs.aws.amazon.com/clean-rooms/latest/userguide/data-formats.html)
+**Related documents:**
+
+  * [AWS WAF Bot Control](https://docs.aws.amazon.com/waf/latest/developerguide/aws-managed-rule-groups-bot.html)
+
+  * [SageMaker AI Managed Spot Training](https://docs.aws.amazon.com/sagemaker/latest/dg/model-managed-spot-training.html)
+
+
+
+
+**Related services:**
+
+  * [AWS WAF](https://aws.amazon.com/waf/)
+
+  * [Amazon ElastiCache](https://aws.amazon.com/elasticache/)
+
+  * [Amazon SageMaker AI](https://aws.amazon.com/sagemaker/)
@@ -37 +90 @@ Use efficient storage formats and streamlined query configurations to reduce unn
-  * [Data Analytics Lens](https://docs.aws.amazon.com/wellarchitected/latest/analytics-lens/best-practice-10.4---partition-your-data-to-avoid-unnecessary-file-reads.html)
+  * [Amazon GuardDuty](https://aws.amazon.com/guardduty/)
@@ -48 +101 @@ To use the Amazon Web Services Documentation, Javascript must be enabled. Please
-Privacy-enhanced collaboration
+Agentic AI cost optimization 
@@ -50 +103 @@ Privacy-enhanced collaboration
-Data transfer and processing
+ADVCOST05-BP02 Select the right foundation model for each advertising agent's task complexity