(WATCH VIDEO) Barnard Computer Science Seminar - L. Jean Camp - Indiana University - May 1, 2023

Speaker Overview and Background

Presenter: Professor Jean Camp, Professor of Informatics and Computer Science at Indiana University.
Host/Introducer: Rebecca Wright, Director of Computer Science at Barnard College.
Academic and Professional Background:
- Double major in Electrical Engineering and Mathematics.
- Master of Science in Engineering ( $\text{MSE}$ ) in Optoelectronics from the University of North Carolina at Charlotte.
- Engineered at the Catawba Nuclear Station early in her career.
- Doctorate from Carnegie Mellon University.
- Senior Member of the Technical Staff at Sandia National Laboratories.
- Eight years on the faculty at Harvard's Kennedy School of Government, with courses cross-listed in Harvard Law School, Harvard Business School, and the Engineering Systems Division at the Massachusetts Institute of Technology ( $\text{MIT}$ ).
- Visiting Professor of Philosophy at Erasmus University in Rotterdam.
- Visiting Scientist at the $\text{MIT}$ Laboratory for Computer Science ( $\text{LCS}$ ).
- Congressional Fellow on Capitol Hill (applied in $2009$ , served in $2010$ , present in the gallery when the Affordable Care Act passed in August $2010$ ).
- Visiting Researcher for Long-Term Cybersecurity ( $2019$ ), a joint position between University of California, Berkeley Electrical Engineering and Computer Sciences ( $\text{EECS}$ ) and Berkeley Law.
Honors and Distinctions:
- Fellow of the Association for Computing Machinery ( $\text{ACM}$ ).
- Fellow of the Institute of Electrical and Electronics Engineers ( $\text{IEEE}$ ).
- Fellow of the American Association for the Advancement of Science ( $\text{AAAS}$ ).
- Member of Sigma Xi, the National Research Honor Society.
- Co-instituted the first caregiver travel scholarship with Rebecca Wright to fund conference attendees bringing caregivers.
- Co-presented at Black Hat with Sanchari Das, where their team constituted $25\%$ of the diversity among Black Hat presenters for that session.
- Developed early technical solutions for distributed voting systems submitted to the Democratic National Convention ( $\text{DNC}$ ).

Interdisciplinary Nature of Network Trust and PKI

Scope of Network Layers:
- Trust spans all $9$ layers of the network model, extending from the physical power grid (e.g., nuclear power station) up to the human operator.
- Without continuous physical power, the physical layer cannot function, rendering higher technical layers useless.
Core Inquiry of Trust Governance:
- Who governs trust across the network?
- Who can modify system configurations, network settings, infrastructure, and implants?
- Who is at risk when trust mechanisms fail? (Failure impacts all network participants).
Traditional vs. Structural Vulnerabilities:
- Traditional security focuses on code-level software bugs, code subversion, and exploits.
- Structural trust governance examines scenarios where software operates perfectly, code is not subverted, and key exchanges succeed, yet system manipulation still occurs due to misplaced trust authority.

Case Study: Russian Cyberattack on Ukrainian Infrastructure

Incident Dynamics:
- During the military invasion of Ukraine, Russian network attackers successfully induced Ukrainian targets to install destructive software originating from an unknown Russian firm.
Cryptographic Validity of the Malware:
- The malware possessed a fully valid digital code-signing certificate issued by DigiCert.
- The certificate was approved by Microsoft and successfully validated up the entire trust chain to the root authority.
- All integrity checks, cryptographic key exchanges, and hash verifications passed without error.
Core Failure Mode:
- The system achieved perfect technical integrity and cryptographic authenticity, but the issuing authority and recipient intent were malicious. The valid certificate authorized software intended to wipe Ukrainian computer systems.

Public Key Infrastructure (PKI) Governance and Architecture

Hierarchical Structure:
- Global Web $\text{PKI}$ authority is concentrated among a small number of Root Program Operators:
  - Google: Dominant operator in the ecosystem.
  - Mozilla: Primarily relies on Google's root store.
  - Apple: Maintains custom key infrastructure for code signing, operating system lists, and Safari browser lists.
  - Microsoft: Operates root lists embedded within its operating system.
  - Tor: Relies on Mozilla, which in turn leverages Google.
Governance Mechanism:
- Certificate Authority Browser Forum (CAB Forum): The primary governing body consisting of root program operators and Certificate Authorities ( $\text{CAs}$ ).
- Root program operators select root $\text{CAs}$ and establish compliance policies.
- Root program operators sign root keys; root keys sign intermediate $\text{CAs}$ ; intermediate $\text{CAs}$ issue end-entity certificates to domain owners.

User Perceptions and Expert Perspectives on PKI

Non-Expert User Survey:
- Sampling Locations: Farmer's markets, shopping malls, fandom conventions (e.g., Gen Con), and academic/industry conferences.
- User Beliefs: Non-experts expressed extreme, unwarranted optimism regarding the assurances provided by a certificate or HTTPS lock icon:
  - Believed a certificate guarantees local legal accountability and consumer protection rights within their jurisdiction.
  - Believed a certificate guarantees strong website privacy policies.
  - Believed a certificate indicates active antivirus scanning and safety from malicious hackers.
  - A significant portion of respondents answered "I don't know."
- Takeaway: Non-experts focus heavily on legal accountability and privacy, possessing no baseline understanding of the technical limits of $\text{PKI}$ .
Expert Interviews and Workshops:
- Venues: Telecommunications Policy Research Conference ( $\text{TPRC}$ ) and Real World Crypto ( $\text{RWC}$ ).
- Shared Expert Consensus:
  - Desire for better code quality and faster revocation responses to rogue certificates.
  - Need for nuanced, contextual, and less frequent security warnings.
  - Unanimous opposition across all disciplines to exceptional access (system backdoors).
- Interdisciplinary Disagreements:
  - Policy experts expected technologists to solve trust failures through technical mechanisms.
  - Technologists expected policy makers to mandate legal and regulatory fixes.

Classification of Certificate Failure Modes

Rogue Certificates:
- Purposeful, false mis-issuances where a certificate is knowingly issued to an unauthorized entity (e.g., issuing an unauthorized certificate for google.com).
- Real-World Example: Healthcare providers issuing validly structured, but completely fraudulent, $\text{COVID-19}$ vaccination cards.
Misused Certificates:
- Certificates issued legitimately, but where the recipient misrepresents identity, authorization, or intent (e.g., typosquatting domains such as g00gle.com).
Passable / Invalid Certificates:
- Certificates that appear legitimate to validation software due to software vulnerabilities, weak cryptographic primitives, or improperly generated local peer $\text{IDs}$ .

Empirical Study of Certificate Incidents (2010–2022)

Dataset Structure:
- Multi-year repository of global certificate data spanning $2010$ to $2022$ , with detailed qualitative coding completed through $2019$ .
Historical Root CA Failures:
- $2010$ marked the start of major public root failures, including the compromise of StartCom (the first major root $\text{CA}$ hack).
- WoSign / StartCom Incident: Following the planned deprecation of the $\text{SHA-1}$ hashing standard, WoSign acquired StartCom and sold backdated $\text{SHA-1}$ certificates to bypass industry deadlines.
2019 Empirical Incident Coding Findings:
- $71\%$ of all certificate incidents were reported by third parties rather than self-reported by the issuing $\text{CA}$ .
- Incidents occurred across highly respected entities, including Swisscom, Belgian root $\text{CAs}$ , and specialized scientific $\text{CAs}$ .
- Primary Causes of Incidents:
  - Non-compliant certificate fields (cryptographic key strength, missing/broken revocation capabilities).
  - Execution of vulnerable software on $\text{CA}$ infrastructure.
  - Flawed business models (selling certificates without identity verification).
  - Ignorance of compliance standards and governing laws.
  - Only $1$ analyzed incident was attributable to a single isolated human operational error.
Concentration vs. Market Share:
- Mis-issuance rates do not correlate directly with market share. High error volumes occurred in medium/small $\text{CAs}$ (e.g., SwissSign, Alma Flora) as well as larger entities (e.g., SSL.com).
Geographic Patterns:
- The United States issues the absolute majority of global certificates.
- Southern European $\text{CAs}$ exhibit a disproportionately high rate of certificate errors relative to their total issuance volume.
Impact of Certificate Transparency (CT):
- In $2019$ , Certificate Transparency became mandatory for all connections in Google Chrome.
- $50\%$ of reported incidents transitioned to third-party detections as $\text{CT}$ logs made non-compliant issuances publicly auditable.

Misconceptions, Market Distortions, and Consumer Awareness

Institutional Over-Promises:
- $2019$ Internet Society ( $\text{ISOC}$ ) publications claimed $\text{PKI}$ provides identity authentication for every user and device on the network, along with complete data/system integrity.
- DigiCert promotional publications claimed $\text{PKI}$ alone serves as the complete foundation for Zero Trust architecture.
Consumer Risk Awareness Disconnect:
- In a classroom survey of $60$ undergraduate cybersecurity students at Indiana University:
  - $1$ student knew if their hardware utilized an Infineon cryptographic chip.
  - $60$ out of $60$ students raised their hands to confirm usage of Steam.
- Demonstrates that consumers are constantly exposed to underlying hardware/certificate risk without functional visibility.
Extended Validation (EV) Failure:
- Extended Validation ( $\text{EV}$ ) certificates were introduced to guarantee verified identity.
- Analysis of $\text{CT}$ data revealed that $\text{EV}$ certificates failed to reduce phishing or mis-issuance rates, functioning primarily as a mechanism for $\text{CAs}$ to charge higher fees.

Certificate Observatory and Machine Learning Classifiers

Certificate Observatory Infrastructure:
- Holds approximately $1\,\text{TB}$ of certificate data.
- Monitors changes in the top $1\,000\,000$ domain certificates weekly (noting the top $100\,000$ domains exhibit almost zero daily key changes).
- Scrapes PhishTank hourly.
- Leverages browser extension volunteers in restricted geographic jurisdictions.
Machine Learning Phishing and Rogue Detection:
- Evaluation Dataset: PhishTank data matched against the top $6$ United States banking institutions (which maintain distinct certificate quality standards) and general web domains.
- Methodology: Standard $10$ -fold cross-validation.
- Architectures Tested:
  - Single $7$ -class classifier.
  - Six distinct binary ( $2$ -category) classifiers (e.g., "IU vs. non-IU", "Bank vs. non-Bank").
  - Binary Decision Tree classifier evaluating Phishing/Rogue vs. Legitimate domains.
- Performance and Rarity Constraints:
  - Decision trees performed best.
  - Rogue certificates are extremely rare in wild data: out of approximately $4 \times 10^6$ analyzed certificates, tens of thousands were phishing, while only $\approx 100$ were true Rogue certificates ( $\approx 0.0025\%$ ).
  - Models trained on academic data were validated against Microsoft internal enterprise datasets (e.g., verifying CertSense detections of Indian $\text{CCA}$ rogue certificates).
- Issuer-Specific ML Classification:
  - Focusing classifiers on specific problematic $\text{CAs}$ increases detection accuracy for those issuers, but fails to predict which $\text{CA}$ will issue the next unknown rogue certificate.

Technical Approaches to Key Pinning and Trust Anchors

Browser/Domain Pinning (2011):
- Hardcoding specific expected certificates inside the browser for explicit domains (e.g., Chrome pinning Google domains).
- Effective for pinned domains, but unscalable across the entire web. Pre-clearing all global certificates introduces severe anti-competitive and legal liabilities for browser vendors.
HTTP Public Key Pinning (HPKP / Operator Pinning):
- Allowed website operators to assert key ownership via HTTP headers.
- Deprecated and discontinued because malicious actors compromised sites and pinned malicious keys, locking out legitimate operators.
DNSSEC / DANE and Service Constraints:
- DANE: DNS-based Authentication of Named Entities.
- Stuxnet Context: The attack on Iranian centrifuges utilized valid software licensing certificates (suffering from hash collisions) rather than code-signing certificates.
- Enforcing service certificate constraints (restricting certificates strictly to designated key usages) was adopted broadly following the Iranian centrifuge disruptions.
Regional Trust Anchors:
- $2014$ European Union proposal to mandate regional trust anchors.
- Flawed approach because European $\text{CAs}$ exhibit the exact same commercial incentives, software vulnerabilities, and error profiles as non-European $\text{CAs}$ .
Third-Party Pinning Authorities:
- Routing domain verification requests through independent entities (e.g., $\text{EFF}$ ).
- Introduces significant user privacy degradation by centralized tracking of browsing behavior.

Household Deployment Study and Visual Contextual Warnings

Field Experiment Setup:
- In late $2019$ , security appliances were shipped to $8$ non-technical households to establish local secure networks.
Dual-Model Detection Pipeline:
- Model 1 (Feature Extractor): Evaluated $\approx 40$ parameters, including:
  - Date of issuance vs. Date first observed on the network.
  - Mismatch between issuer geographic jurisdiction and subject geographic jurisdiction.
  - Indicator Rule: Top $100\,000$ sites roll out certificates far in advance of expiration. A certificate for a top domain observed on its exact issuance date indicates high risk.
  - Jurisdiction Rule: An Indian transportation $\text{CA}$ that historically issued locally suddenly issuing a certificate for Google or Facebook in California triggers a rogue flag.
- Model 2 (Historical Context): Evaluates prior user interaction ("the best model for me today is me yesterday"). Alerts the user upon the first occurrence of an unknown certificate/domain combination.

The Podesta Phishing Attack Case Study

Attack Vector (2016 DNC Incident):
- John Podesta received a malicious email containing a link to a fake Google login page hosted on a top-level country code domain for Tokelau (.tk).
System Failure:
- The phishing domain possessed an active $\text{SSL}$ lock icon and valid certificate logged in $\text{CT}$ .
- Standard browser interfaces presented no warnings because the cryptographic syntax was fully compliant.
Contextual Warning Mitigation:
- A contextual trust interface would explicitly display: "First visit to country code domain .tk, hosted in Region X, signed by a $\text{CA}$ in Country Y with zero prior interaction history."
- Displaying geographic and historical mismatches forces the user to switch from automatic habituated behavior ("fast thinking") to deliberate evaluation ("slow thinking").

Policy and Geopolitical Conclusions

CAB Forum Governance Inadequacy:
- A governance structure designed in the $1990\text{s}$ to secure basic e-commerce credit card transactions is fundamentally incapable of managing modern internet security or international geopolitical conflicts.
Adversarial Machine Learning Limits:
- Detection models achieving $90\%$ to $99\%$ accuracy remain vulnerable because human adversaries deliberately alter techniques to bypass fixed features.
- Unlike biological viruses or physical infrastructure (e.g., bridges), cyber adversaries actively react to and circumvent defense mechanisms.

Questions & Discussion

Question on Geographic Hosting Isolation: How can user risk decisions impact the economics of bulletproof hosting?
- Response: Bulletproof hosting thrives in small jurisdictions that earn revenue by hosting malicious sites while shielding local citizens from harm. If security systems isolate bulletproof-hosted sites so they can only access traffic within their local jurisdiction, the macroeconomic incentive to host malicious infrastructure disappears.
Question on Browser Warning Fatigue and User Decision Making: Do users ignore security warnings regardless of formatting?
- Response: Current browser warnings fail because they focus on technical syntax (e.g., self-signed certificates) and are semantically meaningless to users. Users learn to click through them using fast thinking (Kahneman & Tversky framework). Making warnings visual, contextual (showing maps, location, historical connection data), and rare forces users into slow thinking modes.
Discussion on User Education vs. Systemic Provider Responsibility (Raised by Mary): Should cybersecurity rely on end-user education?
- Response (Mary & Jean Camp): User education wears off quickly and is an intractable solution. Every year brings a new population of $7$ -year-olds getting mobile devices and $70$ -year-olds managing retirement funds (with $15\%$ of the U.S. population being functionally illiterate). Responsibility must shift away from the end user to system-level protections implemented by Internet Service Providers ( $\text{ISPs}$ ) like Comcast and AT&T bundling security directly into residential modems.
Question on Machine Learning Classifier Features: What specific features were most predictive in identifying rogue certificates?
- Response: The most predictive features among the $40$ analyzed were:
  1. The temporal delta between the certificate issuance date and the date it was first observed on the global network.
  2. Geographic jurisdiction mismatches between the issuing Certificate Authority and the subject domain entity.