Sign In

Communications of the ACM

Privacy and security

As Simple As Possible ? But Not More So


EINSTEIN screen depicting volume across U.S. federal government networks by protocol.


The problem: secure the cybernetwork of a large enterprise. It could be one supplying a vast array of services including disease tracking, astrophysics research, weather predictions, and veterans' health care services to a population of 300 million. Or it could be even bigger, perhaps providing critical infrastructure to the population. Requirements include doing so cheaply and efficiently. The solution? The one the U.S. government has begun to implement—EINSTEIN—involves centralizing its Internet connections so as to perform intrusion detection and prevention on all incoming communications.

Because providing services to the public is a fundamental role for U.S. federal civilian agencies, many agencies turned to the Internet in the 1990s. While confidentiality, integrity, and authenticity dominated early federal thinking about Internet security, agencies faced phishing, IP spoofing, botnets, denial-of-service attacks (DoS), and man-in-the-middle attacks. By the early 2000s, the growing number of attacks on U.S. civilian agency systems could not be ignored. The U.S. government's solution has been to build centralized intrusion detection systems (IDS) and intrusion prevention systems (IPS) at large scale. The project, called EINSTEIN, works at an agency-wide, and in some cases, multi-agency-wide level. Federal civilian systems have two million direct users and serve many more.

Few doubt the value of IDS and IPS as part of a cybersecurity solution, but can a centralized system such as EINSTEIN really work? What attacks will EINSTEIN prevent? What will it miss? What are the privacy implications of using the interception program? These are the questions to which we sought answers. Our answers are provisional because few technical details of the system are public, but these answers have become more important in light of the proposed extension of EINSTEIN to critical infrastructure.2 Indeed, as this column went to press, the Washington Post reported that three Internet carriers—AT&T, Verizon, and CenturyLink—were filtering traffic to 15 defense contractors using tools developed by the NSA.4 Of course critical infrastructure should deploy intrusion detection and intrusion prevention systems, but we are skeptical whether the consolidation and real-time information sharing model central to EINSTEIN 3 can effectively migrate to such privately held systems.

Back to Top

EINSTEIN Project Efforts

The purpose of the 2004 EINSTEIN was to do real-time, or near real-time automatic collection, correlation, and analysis of computer intrusion information. IDSs were to be located at federal agency access points to the Internet. If incoming traffic appeared "anomalous," session information would go to US-CERT, the U.S. Computer Emergency Readiness Team, a federal clearinghouse for cyber intrusion information,a But information sharing did not happen in real time and EINSTEIN's voluntary nature meant that many agencies did not participate.

Part of the difficulty was Internet connections. Every small, medium, and large federal agency was connected to the network, sometimes in multiple ways, making control of incoming data and real-time information sharing extremely difficult. The government went about reducing the number of federal connections to the public Internet from a few thousand to a few hundred.

The next program, EINSTEIN 2, uses devices located at the Internet access points to monitor traffic coming into or exiting from government networks and to alert US-CERT whenever traffic matching signatures, patterns of known malware (for example, the IP address of a server known to be hosting malware or an attachment known to include a virus), were observed in incoming packets.7 Participation lagged, but EINSTEIN 2 is now mandatory for federal agencies.

Size is not the only issue in transitioning EINSTEIN systems from federal civilian agencies to the private sector.

The third effort, EINSTEIN 3, will really up the ante by using intrusion prevention systems to stop malware from reaching government sites. EINSTEIN 3 devices will be performing deep packet inspection of content, discarding suspect traffic before it reaches federal systems. (The architecture is such that only communications destined for the federal government are so inspected.) As of this writing, EINSTEIN 3 has been tested only at a single medium-sized federal agency.

Because EINSTEIN IDSs and IPSs would operate on all traffic destined for federal networks, the system would intercept private communications of federal employees (for example, if a federal employee used an agency computer to check a private email account during lunch). But this surveillance is not different from that experienced by employees at regulated industries using company-supplied equipment for their personal communications, and so does seem unreasonable.

Back to Top

EINSTEIN Project Concerns

Can EINSTEIN work? That depends on what "work" means. We have the following concerns.

Scale. Denial-of-service (DoS) attacks can be daunting; they have been measured at 100Gb/s. It is unlikely that the current generation of any network device would be able to resist the DoS attacks at this rate—let alone new attack rates likely in the near future.

Ability to do correlation. Correlation is about discovering previously unknown threats in network traffic in real time as they appear. But this is impossible to do in all but very small networks. No one knows how to use a percentage of the traffic—whether compressed, diarized,b or sampled—to characterize arbitrary new threats. If one is hoping to deter all threats (and not just previously known ones), all incoming data must be correlated and analyzed. That is the crucial point.

One way to think about potential correlation solutions is that architectures can range from highly "centralized" to fully "decentralized" while sensors can be "smart" or "dumb," that is, having the ability to do lots of computation locally, or not.

If analysis is done locally at the data collection point, then the need to see all incoming data requires that all raw signals be sent to all sensors. This quickly becomes unmanageable. If there are n sensors, then each sensor must look at the data from (n−1) other sensors, and there are n(n–1)/2 pairs of data traversing the network. This is simply unmanageable when n is at all large (EINSTEIN is designed to have between one and two hundred collection points). And the sensors would also need protecting.

An alternative approach would be to centralize the data to perform the correlation. Because summarizing the data cannot solve the problem, all the data must travel through the system to the centralized detector. (We note that in an IP-based environment, packet summary information constitutes 2%–30% of the data; the numbers vary depending on whether the packets are carrying email, which comes in large packets, VoIP, which uses small packets, or something in between. In any case, summarizing does not provide savings at anything like the same scale it would for telephone communications.) This is both enormously costly for a network of any scale, as well as unable to provide the millisecond response needed in a serious attack.

(Of course, one could try a middling solution: neither fully decentralized nor fully sharing signals. Depending on where one sets collection, the problems described here will still occur. The two alternative solutions—dumb sensors and decentralized architectures or smart sensors and centralized architectures—have the worst of both worlds: they would either miss the problems, or involve enormous investment. Neither is viable.)

In short, correlation at the scale and speed at which a system serving two million users is expected to operate is not achievable using common production technology.

Device management. Many EINSTEIN devices will be in non-government facilities, but will need to be remotely controlled by US-CERT. Protecting control mechanisms and pathways against intrusion, disruption, modification, and monitoring will be very challenging.

Signature management. EINSTEIN 3 will use classified signatures developed by the government as well as unclassified signatures from commercial IDS and IPS vendors. These signatures will have to be protected from the access point operators as well as from Internet-based attackers.

These complexities make it highly unlikely that EINSTEIN can achieve the job for which it is being designed.

We have concerns about cost. If we assume the IDS/IPS function at a federal civilian agency will be similar to that in commercial network defense products built by Narus, for example, a back-of-the-envelope calculation shows each router directing traffic will require 64 times as much equipment to perform EINSTEIN-type filtering.1 This is clearly a losing battle. In addition, it means that the EINSTEIN program, or at least the instantiation of EINSTEIN 3, will cost approximately $1 billion just for equipment.

EINSTEIN also raises policy concerns. Any IDS looking for long-term subtle attacks must store large amounts of traffic for non-real-time analysis. System design and configuration will determine what is stored and when. The data EINSTEIN collects will have many possible uses. History has shown that investigatory tools are often misused by those with the tools.5,8 There is a significant risk of mission creep for EINSTEIN, and generating detailed logs for all functions that the EINSTEIN 3 device has been configured to do is crucial. Current EINSTEIN 3 documentation does not describe details of the auditing system. Given the size and scope of the EINSTEIN effort, these should be public.

What EINSTEIN can accomplish is limited. EINSTEIN documentation mentions threats of phishing, IP spoofing, botnets, denial-of-service attacks, distributed denial-of-service attacks, man-in-the-middle attacks, or the insertion of other types of malware,6 without noting that EINSTEIN-type systems cannot prevent IP spoofing, man-in-the-middle attacks, and some phishing attacks.

Our bigger concern, however, is in the potential of extending the centralized intrusion-detection/intrusion-prevention to critical infrastructure, including communications and public utilities such as the energy smart grid.9 This is contradictory—a classified U.S. federal government program for protecting widely used private-sector systems. Exacerbating this issue is the fact that the architectures and functions of EINSTEIN and privately held critical infrastructure simply don't match.

Federal civilian systems have two million employees, but critical-infrastructure systems in the U.S. serve over 300 million Americans daily. Can a program that effectively protects the communications of federal agencies with 100,000 employees each do the same for communications giants that instead serve 100 million people? The smart grid, with hundreds of transactions per day to hundreds of millions of endpoints, far exceeds the traffic EINSTEIN is designed to handle.

Size is not the only issue in transitioning EINSTEIN systems from federal civilian agencies to the private sector. While the U.S. government can mandate the types of technologies used by federal agencies, typically the types of systems used in the private sector cannot be so mandated. The biggest problem, however, is the lack of applicability of the EINSTEIN technology to privately held critical infrastructure.

Consider the public communication networks. In the 1990s, the rate of communications transmission was sufficiently slow that the communications bits could be effectively examined and stored—at least if one did sampling. That is no longer true. Meanwhile, communications technologies are in a state of constant innovation. For proper functioning, IDS and IPS should be designed to prohibit those types of communications that are not explicitly allowed. Use of this type of block-first technologies would delay deployment of innovative communications technologies. This would have a devastating impact on U.S. innovation and competitiveness.

Or consider the power grid, which is a loosely coupled federation of many independent (sometimes competing) parties with complex trust relationships.3 This architecture vastly complicates consolidation of the type required by EINSTEIN. Even if consolidation were possible, the need for timely delivery of real-time data and the requirement of high reliability make it undesirable to circuitously direct grid control data through a small number of consolidated access points. In the power grid, function mismatch creates another problem. Centralized IDS/IPS solutions useful for protecting U.S. federal government computer networks may not match well to the power grid. Many parties in the energy grid already have their own IDS/IPS and firewall solutions from a variety of vendors, making the EINSTEIN 3 equipment at least partially redundant. These existing IDS/IPS solutions are often integrated with other important functionality such as quality-of-service, compression, and SCADAc reports (which are part of Critical Infrastructure Protection requirements for the North American and Federal Energy Regulatory Commission). While these reports are generated by the same equipment that performs IDS and IPS, EINSTEIN 3 equipment cannot realistically subsume this functionality.

It is far from clear that this billion-dollar system can deliver sufficient security to be worth the cost.

Putting it simply, there are deep and fundamental differences between communication networks supporting the U.S. federal government and those supporting private sector critical infrastructure. These differences create serious problems in any attempt to extend EINSTEIN-type technologies to private-sector systems controlling critical infrastructure. This is true in the U.S. and, depending on architecture, may be true elsewhere.

Back to Top


Even implementing EINSTEIN in the restricted environment of federal civilian agency systems is highly complex, and it is far from clear that this billion-dollar system can deliver sufficient security to be worth the cost. In the domain of privately owned critical infrastructure, the potential of EINSTEIN is much less clear. Electronic fences protecting critical infrastructure sound good, but once one examines network architecture more carefully, EINSTEIN's fit is highly questionable. In determining how to protect critical infrastructure, one should keep in mind what Einstein himself was purported to have said: "Everything should be made as simple as possible, but no simpler"—and then develop solutions accordingly.

Back to Top


1. Bellovin, S.M., Bradner, S.O., Diffie, W., Landau, S., and Rexford, J. Can It Really Work? Problems with Extending EINSTEIN 3 to Critical Infrastructure. Harvard National Security Journal, to be published.

2. Hoover, N. Cyber command director: U.S. needs to secure critical infrastructure. Information Week (Sept. 23, 2010);

3. Juniper Networks. Smart Grid Security Solution: Comprehensive Network-Based Security for Smart Grid, 2010.

4. Nakashima, E. NSA allies with Internet carriers to thwart cyber attacks against defense firms. Washington Post (June 16, 2011).

5. United States Congress. Senate, Select Committee to Study Governmental Operations with Respect to Intelligence Activities, Final Report of the Select Committee to Study Governmental Operations with Respect to Intelligence Activities: Supplementary detailed Staff Reports on Intelligence Activities and the Rights of Americans: Book II, Report 94–755, 1976.

6. U.S. Department of Homeland Security, Computer Emergency Readiness team (US-CERT). Privacy Impact Assessment for the Initiative Three Exercise. Washington D.C., 2010.

7. U.S. Department of Homeland Security, Chief Privacy Officer. Privacy Impact Assessment for EINSTEIN 2. Washington D.C., 2008.

8. U.S. Department of Justice, Office of the Inspector General (Mar. 2008). A Review of the FBI's Use of National Security Letters: Assessment of Corrective Actions and Examination of NSL Usage in 2006.

9. Zetter, K. Let us secure your network for you or face the "wild, wild west" Internet alone. WIRED (may 27, 2010.)

Back to Top


Steven M. Bellovin ( is a professor in the computer science department at Columbia University.

Scott O. Bradner ( is University Technology Security Officer at Harvard University.

Whitfield Diffie ( is Vice President for Information Security and Cryptography at the Internet Corporation for Assigned Names and Numbers (ICANN).

Susan Landau ( is currently a fellow at the Radcliffe Institute for Advanced Study at Harvard University.

Jennifer Rexford ( is a professor in the computer science department at Princeton University in New Jersey.

Back to Top


a. US-CERT collects information from federal agencies, industry, the research community, state and local governments, and sends out alerts about known malware; see

b. "Diarize" is used by the trade to mean making a diary of the data; in the case of a telephone call, this might be the to/from, time, and length of the call, while for IP communications, this would be the metadata of source and destination IP addresses, TCP source and destination ports, and perhaps length of packet.

c. SCADA (Supervisory Control And Data Acquisition) systems are used to monitor and control industrial processes.

Back to Top


UF1Figure. EINSTEIN screen depicting volume across U.S. federal government networks by protocol.

Back to top

Copyright held by author.

The Digital Library is published by the Association for Computing Machinery. Copyright © 2011 ACM, Inc.


No entries found