Skip to main content

Fairness, Privacy, and Harm in Networks

Network data describes people in relation to others. A decision about one node can therefore expose, rank, or constrain neighbors who never supplied data or consented to the use. Removing names is not enough: topology and auxiliary information can support reidentification.

Prerequisites​

Complete graph evaluation and leakage. A leakage-safe score can still be harmful, unlawful, unrepresentative, or inappropriate for the decision.

Learning goals​

  • Separate observation bias, representation harm, allocation harm, and privacy risk.
  • Explain why topology can reidentify nominally anonymous nodes.
  • Define affected people beyond labeled model subjects.
  • Recognize when group-disparity claims are unsupported by the available data.

Harm-analysis contract​

  • Affected parties: represented people, their neighbors, people missing from the graph, moderators/operators, and communities described by aggregate claims.
  • Sensitive surface: node attributes, edges, timestamps, embeddings, neighborhoods, predictions, explanations, and query access.
  • Evidence: structural signatures and schema coverage are screening signals, not a privacy guarantee or fairness certification.
  • Interpretation limit: no protected-group attributes or stakeholder evidence exist in the fixture, so group fairness cannot be measured here.

Trace harm through the pipeline​

Collection can overobserve highly active people and underobserve private, offline, or marginalized relationships. Graph construction can encode institutional categories as natural facts. Embeddings can place structurally isolated people far from a majority neighborhood and reproduce sensitive information even when that attribute is omitted. Ranking or link prediction can then allocate attention, investigation, credit, safety interventions, or exclusion.

Representation harm includes stereotyping, erasure, denigration, and treating a contested group boundary as objective. Allocation harm changes opportunities or resources. Both require domain and stakeholder evidence; parity of one model metric cannot establish acceptability.

Privacy review must consider node-level and edge-level disclosure, attribute inference, membership inference, reidentification with auxiliary graphs, and harms to nonusers connected to users. Controls include purpose limitation, data minimization, access logging, aggregation thresholds, retention limits, red-team queries, and a documented deletion/retraining path. Formal privacy mechanisms require a threat model and privacy budget; deleting names is not such a mechanism.

Worked example​

import csv
from collections import Counter
from pathlib import Path

fixture = Path("tests/fixtures/graph-data-science")
with (fixture / "interaction-nodes.csv").open(encoding="utf-8") as handle:
nodes = list(csv.DictReader(handle))
with (fixture / "interaction-relationships.csv").open(encoding="utf-8") as handle:
edges = [row for row in csv.DictReader(handle) if row["record_status"] == "valid"]

signatures = Counter()
for node in nodes:
node_id = node["node_id"]
signature = (
sum(row["source"] == node_id for row in edges),
sum(row["target"] == node_id for row in edges),
sum(row["source"] == row["target"] == node_id for row in edges),
)
signatures[signature] += 1
unique = sum(count == 1 for count in signatures.values())
has_group_attributes = any(column in nodes[0] for column in ("protected_group", "affected_group"))
print(f"structural_signatures={dict(sorted(signatures.items()))} unique_signature_classes={unique}")
print(f"affected_group_attributes_available={has_group_attributes} disparity_audit=unsupported")
print("privacy_boundary=pseudonyms-do-not-prevent-topology-based-reidentification")

Expected output:

structural_signatures={(0, 0, 0): 1, (1, 1, 0): 1, (1, 2, 0): 1, (2, 1, 0): 1, (2, 2, 0): 2, (2, 2, 1): 1} unique_signature_classes=5
affected_group_attributes_available=False disparity_audit=unsupported
privacy_boundary=pseudonyms-do-not-prevent-topology-based-reidentification

Uniqueness in this tiny fixture is not an attack success estimate. It only demonstrates that structural patterns can distinguish pseudonymous records.

Limitations​

  • Protected attributes can be unsafe or illegal to collect; absence also prevents quantitative disparity estimates.
  • Inferred demographic labels are error-prone and can create new harm.
  • Network dependence complicates individual-level consent, deletion, and uncertainty.
  • Technical metrics cannot replace affected-group participation, legal review, appeal, remedy, and a decision not to deploy.

Practice​

Write a harm register for a “people you may know” feature. Include one nonuser, one falsely suggested link, one sensitive true link, an owner, a prevention control, a detection signal, and a remedy.

Next steps​

Continue to the local interaction-network project to produce auditable evidence while keeping these limits visible.

Primary sources​

Checked 2026-10-18: A. Narayanan and V. Shmatikov, De-anonymizing Social Networks; NIST, Towards a Standard for Identifying and Managing Bias in Artificial Intelligence; NIST, Privacy Framework.